Rp-1146-Building Operation and Dynamic Within An Aggregated Load PDF

Shaping Tomorrow’s
Built Environment Today

©2012 ASHRAE www.ashrae.org. This material may not be copied nor distributed in either paper or digital form without
ASHRAE’s permission. Requests for this report should be directed to the ASHRAE Manager of Research and Technical
Services.
ASHRAE Research Project 1146-RP
Building Operation and Dynamics

Within an Aggregated Load
Presented to:
American Society of Heating, Refrigeration and Air-Conditioning Engineers

1792 Tullie Circle, NE
Atlanta, GA 30329
Prepared by:
T. Agami Reddy Civil & Architectural Engineering, Drexel University

3141 Chestnut Street, Philadelphia, PA 19104
and
Leslie K. Norford Tabors Caramanis & Associates

50 Church Street, Cambridge, MA 02138
August 30, 2002
I
COPYRIGHT
AMERICAN SOCIETY OF HEATING. REFRIGERATING
AND AIR-CONDITIONING ENGINEERS. INC. '
1791 TULUE CIRCLE. ATLANTA, 6A 30329
EXECUTIVE SUMMARY
This report documents work performed for ASHRAE 1146-RP, entitled "Building
Operation and Dynamics Within an Aggregated Load." Research was done in two
phases. Tasks undertaken as part of Phase I involved (1) an extended background and
historical perspective on electric utility aggregation, (2) a description of various candidate
sites where load aggregation programs have been implemented and clearly recorded, (3) a
detailed description of each of three case-study sites where load aggregation and
subsequent energy management strategies were performed, and (4) a summary of the
lessons learned from the above case studies.
The case studies revealed that the motivation for aggregating multiple buildings under the
current state of affairs was essentially energy-cost related (i.e. better rates could be
negotiated because of the larger aggregated load). Very little was being done in terms of
load control in either individual buildings or in aggregated buildings.
The second phase of the work consisted of the development of a list of 11 tools that could
be used by aggregators and their customers to alter the operation of buildings to control
load. The tools, separated into three categories, are as follows:
Aggregation of Load-Control Customers: This set of four tools is proposed to aid an

aggregator in assembling a portfolio of customers. To this end, the tools are intended to
minimize calculations that require extensive building-specific information.
• Tool 1: Customer pre-screening: to be able to pre-screen a customer for inclusion
as a viable candidate for an aggregation portfolio
• Tool 2: Assessment of customer load-reduction potential: to provide preliminary
and quick estimates of load reduction potential via lighting reduction, space
temperature and HVAC set point control
• Tool 3: Portfolio optimization: to determine which of the viable customers should
be aggregated to maximize savings and to perform a risk analysis, i.e., analyze the
soundness of the decision in the face of uncertainties and variability in the
available data
• Tool 4: Contractual: to evaluate different types of contractual rates and clauses
between aggregators, customers and load-serving entities
Pro-active load Control: The objective of this set of four tools is to determine how
much load to curtail, and which specific end-use load curtailment measures to implement
and in which specific buildings within the aggregate:
• Tool 5: Load forecasting at several levels: end-use loads for individual customers,
whole-building loads for individual customers, and aggregated loads, as well as
the associated uncertainties
• Tool 6: Thermal-comfort penalty formulation based on the premise that thermal
comfort be quantified at the zone, building and site levels, via formation of a
weighted sum of indices computed at the lowest level.
1
• Tool 7: Interaction with load-serving entities concerning price forecasts that can
be obtained from power pools or from independent organizations
• Tool 8: Load optimization in order to find the optimal combination of load
curtailment measures to implement that can meet the required load reduction
while minimizing the thermal-comfort penalty function
Management of load, power and financial data: This last set of three tools concerns
communication of price and load information, measurement of load-control action, and
financial payments and penalties for load control.
• Tool 9: Communications between an aggregator and its customers, or between a
power marketer and the aggregator's customers who are served by the power
marketer
• Tool 10: Monitoring and verification to determine which customer has reduced
his load, and by how much, after being notified.
• Tool 11: Financial remuneration for allocating the benefits of effective load
control and penalizing ineffective action.
The report describes the purpose of each tool. For some tools, it is noted that previous
work is adequate, as with load forecasting. For others, notably load-reduction potential
and load optimization, the report includes extensively developed examples of
methodology and results. The proposed tool for thermal-comfort penalties features a
method for summing thermal-comfort penalties across zones and buildings. The scope of
this research was not to develop the tools to the extent that concerned load aggregators
can use them. Rather, this research describes the functionality of these tools in a
conceptual manner and provides relevant review of prior work and what future
developments are needed for each of these tools. Due to economies of scale, the scope is
limited to larger commercial customers and not residential and small commercial
customers.
The description of the tools is followed by an illustrative case study, based on DOE-2
simulations, which is meant to demonstrate by means of an illustrative case study the
benefits in multi-building load aggregation. More specifically, the issue of what are the
added financial benefits (and under what circumstances) for each individual building
owner to (i) aggregate with others, and (ii) intentionally curtail his load during peak days
in conjunction with the aggregate of buildings are discussed. This report culminates in a
conclusion chapter that summarizes the report and identifies areas for future work.
2
ACKNOWLEDGMENTS
The authors gratefully acknowledge the assistance of the following individuals during the
course of this research project:
Jason Lukes, graduate student at Drexel University, Philadelphia

Travis Peyton, Ballinger, Philadelphia
Richard Heimann and Itzhak Maor, PWI Energy, Philadelphia
Bill Taylor, Drexel University, Philadelphia
Randolph Haines, Thomas Jefferson University, Philadelphia
Wayne Stowe, American Home Products
Scott Englander, Ezra Hausman and Rick Hornby, Tabors Caramanis & Associates,
Boston
The authors also express their thanks to the ASHRAE Project Monitoring Sub-
committee, especially the Chair John House, for their dedication, helpful suggestions and
guidance during the execution of this project. This research was proposed by ASHRAE
Technical Committee 4.6, Building Operation Dynamics.
3
TABLE OF CONTENTS
EXECUTIVE SUMMARY 1
ACKNOWLEDGMENTS 3
1. OBJECTIVES AND SCOPE 8
2. BACKGROUND 9
2.1 Electric Utility Aggregation 9
2.1.1 Aggregate Billing 9
2.1.2 Load Control 10
2.1.3 Building Optimization - Individual Buildings 10
2.1.4 Load Aggregation-Multiple Buildings 10
2.2 Historical Perspective 11
2.2.1 Participation in Deregulation 11
2.2.2 Restructuring & Retail Access 11
2.2.3 Current Picture of Electric Utility Market 13
2.2.4 Rate Components and Metering 14
2.2.5 Electric Use in Demand Metered Buildings 17
2.2.6 Energy Service Providers and Their Customers 17
3. CANDIDATE SITE SELECTION 21
3.1 General Criteria for Selection 21
3.2 Selected Candidate Sites 21
3.3 Preliminary Site Interviews 22
3.3.1 University of Delaware 22
3.3.2 Wyeth-Ayerst Pharmaceuticals 25
3.3.3 Fort Dodge Animal Research Facility 27
4. CASE STUDY EVALUATION 30
4.1 Objectives 30
4.2 Documentation Methodology 31
4.3 Presentation Format 31
4.3.1 Metered Group Details 32
4.3.2 Building Locator Map 33
4.3.3 Building Summary 33
4.3.4 Metered Group Use Breakdown 34
4.3.5 Chart of Metered Group Use 34
5. DREXEL UNIVERSITY 35
5.1 Site Characteristics 35
5.2 Rate Structure, Utility Profile and Metering 37
5.3 Metered Group Summary 43
5.4 Building Operations and Dynamics 53
5.5 Financial Implications of Aggregation 53
6. THOMAS JEFFERSON UNIVERSITY 56
6.1 Site Characteristics 56
6.2 Rate Structure, Utility Profile and Metering 58
6.3 Metered Group Summary ...62
6.4 Building Operations and Dynamics 70
4
6.5 Financial Implications of Aggregation 70
7. AMERICAN HOME PRODUCTS 72
7.1 Background 72
7.2 Information Exchange 73
7.3 Bid Assessment 74
7.4 Conclusions 76
8. CASE STUDY SUMMARY 77
9. OVERALL DESCRIPTION OF TOOLS 78
9.1 Background 78
9.2 Current and Future Scenarios 78
9.3 Objectives and scope 82
9.4 Simple Example of Benefits of Multi-Building Aggregation 82
9.5 Issues Pertinent to Building Owners and Load Aggregators 83
9.5.1 The Building owner's perspective 83
9.5.2 From the perspective of the pro-active load aggregators 83
9.6 Tools Needed by Local Aggregators Offering Pro-Active Load Management
Services 84
10. TOOLS TO DEVELOP AN AGGREGATION PORTFOLIO (TOOLS 1 - 4 ) 88
10.1 Tool 1: Customer Pre-screening 88
10.1.1 Objective 88
10.1.2 Analytical Criteria 88
10.1.3 Other requirements (metering and current rates) 89
10.2. Tool 2: Assessing Customer Load Reduction Potential 90
10.2.1 Objective 90
10.2.2 Different Methods of Reducing Loads 90
10.2.3 Approaches to Determining Load-Reduction Potential 92
10.2.4. Examples of Estimates of Load Reduction Potential 93
10.3 Tool 3: Portfolio Optimization 109
10.3.1 Objective 109
10.3.2 Mathematical Basis 109
10.3.3 Data Requirements 110
10.3.4 Risk Analysis 110
10.3.5 Illustrative Example ." 110
10.4 Tool 4: Contracts between Aggregators, Customers, and Power Marketers 114
10.4.1 Distinguishing aggregators and load-serving entities 114
10.4.2 Independent marketer 115
10.4.2 Standard and default offers from investor-owned utilities 116
10.4.3 Municipal utility 117
10.4.4 IOU-affiliated marketer 118
10.4.5 Contract enhancements to facilitate load control 118
11. LOAD FORECASTING (TOOL 5) 119
11.1 Background and Definitions 119
11.2 Short-term forecasts of power-system loads 121
11.3 Forecasts of building electrical loads 123
11.3.1 Overview 123
11.3.2 Forecasts Based on Load Research Data 124
5
11.3.3 Whole building electric load models 125
11.3.4 Disaggregated end-use electric models 125
11.3.4 Building thermal load models 126
11.5.1 Electric power consumed by building air handling units 130
11.5.2 Power consumed by pump(s) 130
11.5.3 Power consumed by chiller(s) 132
12. COMFORT PENALTY FRAMING (TOOL 6) 136
12.1 Objective and Scope 136
12.2 Literature Review 136
12.4 Discussion 141
12.5 Summary 143
13. INTERACTION OF AGGREGATORS AND CUSTOMERS WITH POWER
MARKETS AND INDEPENDENT SYSTEM OPERATORS (TOOL 7) 144
13.1 Price forecasts 144
13.1.1 Overview 144
13.1.2 Forecasting Sources 145
13.2 Bilateral contracts in the forward market 148
13.3 Purchases in the spot market 148
13.4 Purchases of reserve power as required by the Independent System Operator... 149
14. LOAD CONTROL VIA GENERATION OF ALTERNATIVE CONTROL
STRATEGIES AND SEARCH FOR OPTIMUM (TOOL 8) 150
14.1 Introduction 150
14.2. Simulation Model 151
14.3. Simulation Protocol and Sample Results 151
14.3.1. Single-Building Optimization 151
14.3.2. Two-Building Optimization 154
14.4. Discussion 156
14.4.1 Lessons from control of lights and air conditioning in a single building and in
two buildings 156
14.4.2 Additional control options 158
14.4.3 Alternative simulation environments 159
15. POST-CONTRACTUAL CAPABILITIES (TOOLS 9 - 11) 161
15.1 Tool 9: Communication Between Aggregator and Customer...." 161
15.1.1 Functionality of tool 161
15.1.2 Literature review 161
15.1.3 Conclusion/Future 162
15.2 Tool 10: Monitoring and Verification 163
15.2.1 Objectives and capabilities 163
15.2.2 Literature review 163
15.2.3 Issues in measurement and analysis 164
15.2.4 Summary of salient features of Tool 10 165
15.3 Tool 11: Financial Remuneration Determination 167
15.3.1 Functionality 167
15.3.2 Future work 167
16. ILLUSTRATIVE CASE STUDY 169
16.1 Objectives 169
6
16.2 Methodology 169
17. CONCLUSION 181
REFERENCES 184
APPENDIX A OPTIMIZATION TOOL 195
A.l Simulation Model 195
A.l.l Model 195
A.1.2 Set of Simulation Cases 196
A.1.3 Simulation Structure 196
A.2 Extensions to the simulation files 200
A.2.1 Base Case 202
A.2.2 Calculation of energy used by lights and equipment 204
A.2.3 Hourly energy consumption during load-control hours 204
A.2.4 Including room-air temperature set point as a control variable 205
A.2.5 Modeling a VAV system 206
A.3 Operation of the simulation environment 208
A.3.1 Switching between an exhaustive search and a GA search 208
A.3.2 Simulation output 209
A.4. Single-Building Optimization 209
A.4.1. Single-case simulations, with electricity used by air conditioning, lights and
equipment 210
A.4.2. Exhaustive search of all possible cases 216
A.4.3. Genetic-algorithm search 225
A.4.4. Use of simulation environment that includes a compilation of hourly energy
consumption during load-control hours 235
A.4.5. Exhaustive search with peak demand included 236
A.4.6. Genetic-algorithm search with peak demand included 236
A.5. Two-Building Optimization 239
A.5.1 Possible Combinations of Control Strategies for Two Buildings 239
A.5.2 Simulation Results 240
7
1. OBJECTIVES AND SCOPE
The overall objectives of ASHRAE 1146-RP were:
(1) to identify situations and conditions under which aggregating individual building
loads is attractive for managing total, multi-building load,
(2) to identify and evaluate operating and control strategies for use in individual
buildings that will reduce energy operating costs at the aggregate level by taking
advantage of the diversity in demand among buildings, and
(3) to develop specific recommendations as to
(a) what engineering developments are needed in accurate prediction of utility load
profiles,
(b) what types of loads in the buildings can be aggregated (i.e., how can the different
loads be classified in a manner suitable for load management),
(c) what load analysis tools are available,
(d) what types of building system control strategies can be used,
(e) what types of control and communication strategies and systems should be
considered (for example EMCS system, internet), and
(f) what types of skill and education needs to be provided to building owners and/or
energy managers, Energy Service Companies (ESCO's) and occupants in the
current utility deregulation atmosphere.
To accomplish these goals, the following discrete tasks were proposed:
Phase 1: Identification of Load Aggregation Opportunities

Task 1.1 Background Information Assembly
Task 1.2 Case Study Selection
Task 1.3 Case Studies
Task 1.4 Preliminary Assessment
Phase 2: Load Aggregation Opportunity Assessment

Task 2.1 Load Aggregation Opportunities
Task 2.2 Define Assessment Approach
Task 2.3 Assessment of Load Aggregation Situations
Task 2.4 Research and Technology Requirements
Task 2.5 Reporting
As will be noted in Chapter 9, the second phase evolved, under the guidance of the
Project Monitoring Subcommittee (PMSC), into a proposal of tools for use by
aggregators rather than an analysis of load-control efforts. This change was motivated by
the lack, at the time of the research, of documented load-control efforts at the level of
aggregates of buildings.
8
2. BACKGROUND
First, this section will briefly describe retail access and the standard electric offer as part
of electric deregulation. Next, the status of commercial and industrial building metering
as it relates to electric rates will be addressed. Finally, some of the recently announced
deals made by competitive service providers will be reviewed and the types of businesses
availing themselves of such service from these providers will be summarized.
2.1 Electric Utility Aggregation
The structure of the electric utility market and the complex relationship between suppliers
and consumers necessitates clarification of some terms contained in this report. The
following sections will distinguish between aggregate billing, load control, building
optimization, and load aggregation.
2.1.1 Aggregate Billing
This pertains to the case when a customer receives one monthly bill for the coincident
peak demand of all his electric meters as compared to a bill that reflects the sum of all the
individual, non-coincident peaks. This type of aggregation takes place at the facility
level, usually a large institution, where it is feasible to combine all existing electrical
services into one large service. A significant load (50,000,000 kWh or more annually)
and adequate load factor (.7 or above) are essential prerequisites for the customer to enter
into negotiations with the electric utility. In some cases, large customers have had to
threaten their electric providers with cogeneration plants before the utilities responded
with alternatives for billing. Aggregate billing, also known as conjunctive billing, is one
of the simplest forms of aggregation, and the one most widely practiced prior to
deregulation. Typically, this agreement has little or no effect on how the individual
buildings are operated or controlled.
Under an aggregate billing agreement, all the building loads are summed together and a
peak is recorded for the highest monthly load. This one-time peak can never be more
than the sum of the individual peaks, but there are instances where it is much less.
Without any changes to his operational schedule or individual building control, an energy
manager can save money just by changing his billing structure.
An alternative would be to request that the energy distributor provide a single power
substation, through which the site would then draw all of its power. In this case, the
distributor bills according to a single, high-voltage service. This often involves the
purchase of the utility substation, but gives the customer savings similar to the scenario
described above.
9
2.1.2 Load Control
This term refers to changes in the operation and dynamics of individual buildings, which
come from shifting load from high-priced hours to low-priced hours. While energy
saving measures may also be employed to reduce consumption, they cannot be
considered to be load control measures unless the customer has the ability to curtail load
in response to some demand signal. The time frame of such a response is a critical issue.
For instance, the installation of energy efficient electronic ballasts in fluorescent lighting
is a logical measure for reducing electric consumption. However, for lighting to be
considered a form of load control, dimmers, occupancy sensors or non-essential overrides
must be provided so that the building operator can curtail demand load as needed. Such
measures will have an immediate effect on the demand load of the building.
In the case of a manufacturing or laboratory facility, this type of immediate response may
not be possible. Because lighting makes up such a small portion of their overall demand
load, these building types will not be able to show a significant decrease by lighting
control measures alone. A manufacturer may have to shut down process lines to show a
sizeable change in demand, and a laboratory conceivably might have to reduce or even
turn off the air-handling units that supply 100% outside air to clinical rooms. Because
manufacturing operations are crucial to the financial well being of the company and
laboratory facilities must meet stringent guidelines for ventilation and indoor air quality,
neither of these measures would be feasible as forms of load control.
2.1.3 Building Optimization - Individual Buildings
Essentially, optimization means ensuring that the load control measures for the existing
equipment will provide the best possible performance. These ambiguous terms,
optimization and performance, depend largely on where the attention of the building
operator is focused. If the customer has a demand-based electric utility rate structure,
optimal performance is likely to mean providing adequate comfort for occupants while
reducing demand costs wherever possible. For our purposes, the term building
optimization will refer to changes in load control strategies that reduce the electric energy
and demand costs to the customer. Load control measures will depend largely on the
type of building and the electric utility rate structure.
2.1.4 Load Aggregation - Multiple Buildings
Once individual buildings have been optimized, they will be grouped together to form
larger blocks of electric power usage. By taking advantage of the diversity in demand
between these multi-building groups, it is possible to reduce the total energy costs even
further than those of the already optimized buildings. The techniques used to optimize
building operation in aggregate are vital to the success of this project. For this reason, we
10
clarify that load aggregation can occur in a variety of forms, but the strategies for multi-
building analysis will be uniform. Also, in these large blocks, groups of buildings can
offer load aggregators a much more attractive load. By guaranteeing that they will
consume a significant portion of the suppliers' available power, such aggregated groups
of buildings will be offered more attractive rates by the electric supplier.
2.2 Historical Perspective
This section explains the conditions prevalent in existing deregulated electric utility
markets, primarily in Massachusetts and Pennsylvania, which give customers a choice of
electricity suppliers. Because building load control is more attractive under demand-
based rate structures, the focus is on the characteristics of these rates and the behavior of
customers who must respond to them.
2.2.1 Participation in Deregulation
While many of the players in the competitive electric market are larger commercial and
industrial campuses, smaller customers also have the option to take advantage of
unbundled prices and better rates for electric usage. The problem is that many small
customers must be aggregated for the supplier to ensure the demand of a significant
portion of his generating capacity. From the supply side, this involves great risk, and
therefore, incentives for such activity are presently minimal. The following sections
describe the process of deregulation and break down customers who take advantage of it
in more detail.
2.2.2 Restructuring & Retail Access
Electric industry restructuring has introduced retail access. Retail access enables end-use
customers to "shop" for the electric services that suit their needs. Such services include
both the energy component (the electricity itself) as well as other "value added" options.
Such options may include billing choices, energy pricing options, and various support
services. Examples of these are:
(1) Consolidated billing - electricity use and demand from multiple meters are combined
on one bill,
(2) Budget billing - billed same amount each month with minor corrections annually
(available for some time),
(3) Fixed price energy,
(4) Energy priced at a benchmark plus or minus some percentage,
(5) Energy use analyses, and
(6) Energy efficiency engineering and design services.
So far, retail access has been implemented in California, Massachusetts, and
11
Pennsylvania. Additionally, restructuring legislation has been enacted in eighteen states,
including Arizona, Arkansas, Connecticut, Delaware, Illinois, Maine Maryland, Montana,
Nevada, New Hampshire, New Jersey, New Mexico, Ohio, Oklahoma, Oregon, Rhode
Island, Texas and Virginia (EIA, 1998). Most remaining states are investigating the
possibility of electric restructuring. In December 1999, the Federal Energy Regulator
Commission ordered all electric utilities to file plans to be part of Regional Transmission
Groups by late 2000. This Order, Order 2000, likely foreshadows further changes in the
industry, which may fuel further consumer options for energy services.
In states where electric retail competition has been implemented, building owners have
three major categories of opportunities to reduce their annual electricity bill - changing
suppliers (supply choice), shifting consumption from high-price hours to lower-price
hours (load control) and reducing the level of consumption (efficiency improvement). Of
these, changing suppliers is the newest option and involves the least capital intensive.
The remaining two options, load control and efficiency improvement, have been
available for many years and often require major capital investments.
To assist customers who elect not to choose an energy supplier, states have created
"default service." Default service, also called the standard offer, is a price-regulated
service offered by providers meeting state-set criteria. The method used to set prices for
standard offer service varies from state to state. For example, in Pennsylvania prices
were set for each customer class, by utility, to approximate retail levels. In contrast, in
Massachusetts a single price was set for all customer classes within each utility. The
Massachusetts prices were set at or below wholesale levels in the initial years and are
scheduled to increase to estimated retail levels over a 7-year transition period.
The terms and conditions of "standard offer" service have a major impact on the benefits
that may be accrued through the use of the energy control strategies mentioned above. In
some utility service territories, the savings due to load control may be lower during post-
deregulation relative to pre-deregulation because of the design of standard offer service
and the policy governing the recovery of a utility's stranded costs. In some instances,
standard offer service prices are so low as to discourage competition from other energy
suppliers because the margins available to them do not exceed the' cost of customer
acquisition. Similarly, some standard service offerings are attractive to customers and the
opportunity cost of researching options may appear to outweigh the savings available—
hence the consumer is not encouraged to leave the default service.
Building owners in states with retail competition have begun to switch from "standard
offer" service to competitive energy suppliers. Suppliers are attracting building owners
with attractive pricing and other service options. The attractive pricing is available
because if a supplier can guarantee sufficient load over a pre-specified duration, he can
frequently obtain power at lower rates than could a building owner acting alone. Such
suppliers may be able to offer electricity at a price less than "standard offer" because their
participants have entered multi-year purchasing commitments and not because they have
a load profile materially different from the load profile of customers taking standard offer
service. Power suppliers focusing primarily on buildings with a load profile materially
12
different from standard offer service appear to be the exception rather than the rule.
Building owners in these situations benefit because the savings on the energy component
can outweigh benefits that may be achieved through load control or efficiency
improvements. This is due in part to how electric prices are structured.
2.2.3 Current Picture of Electric Utility Market
After many discussions with customers and contacts in the electric utility industry, we
developed the following two figures to illustrate our understanding of the market
structure before and after deregulation. These diagrams are modeled after the structure of
Pennsylvania's system, but generally apply to situations in other locations as well.
The first diagram, Figure 2.1, shows how generation, transmission and distribution were
previously bundled into one entity referred to as the utility company. Other components
include independent generators, who sell power to the regional grid or pool, but are not
involved in other sectors, and the end-use customers who are billed by the utility
company.
Power Generating Station Independent Generator
>r V
*w
^
REGIONAL POWER GRID - TRANSMISSION
>f
LOCAL DISTRIBUTION UTILITY
i ^f >f
Consumer A Consumer B
Figure 2.1. Regulated electric utility market structure.
13
The second diagram, Figure 2.2, shows how the components of the utility company
are currently separated, and introduces a new player, the Energy Service Provider
(ESP), who interacts with both suppliers and consumers. It is the ESP who functions
as an active load aggregator, assembling a portfolio of customers into a large block of
load. In some cases, local distributors may withdraw their metering system if a
consumer chooses another power supplier. Under these circumstances, the ESP may
have an independent metering company step in. Also, it should be noted that under
this deregulated system, an Independent System Operator (ISO) manages the
exchanges between suppliers and local distribution. The current system offers an
advantage to customers in the form of retail access as described in Section 2.2.2,
Restructuring and Retail Access.
Power Supplier A , Power Supplier B
Energy Service Provider (ESP)

(Load Aggregator)
>f V
REGIONAL POWER GRID - TRANSMISSION - INDEPENDENT

>f
Independent Metering Company
LOCAL DISTRIBUTION ->
(Acting for ESP)
. " >f >f
Consumer A Consumer B Consumer C
Figure 2.2. Electric utility market structure following deregulation.
2.2.4 Rate Components and Metering
The price for a kWh of electricity is comprised of an energy (commodity) charge, a

transmission and distribution charge, and an "uplift" charge (to pay for stranded costs or
other administrative functions of the energy supplier). Uplift does not include a
marketers' markup. Typically the energy component is 50% or more of the price for a
kWh of electricity. Thus a reduction in the commodity price can impact the end use price
in a noticeable manner.
Many competitive power suppliers offer their customers the opportunity to achieve
additional savings through load control and/or efficiency improvements. However
suppliers provide, and charge for, those services separately from their electricity supply
14
service.
In addition to the kWh charge, commercial and industrial customers historically have
paid some type of capacity (demand) charge and may have been under a special tariff
such as a real-time pricing or bulk purchase rate.
One of the prerequisites for obtaining service under a demand-measured

industrial/commercial rate is a meter that can measure and record demand. The low
penetration of real-time meters and meters with remote communication ability hampers
the development of innovative pricing options. The 1995 EIA Commercial Building
Energy Consumption Survey (CBECS), Table 15, found that 51% (2223/4344) of
buildings surveyed paid a demand charge (Olsen 1998). The information contained in the
CBECS tables in columns with the heading "Demand Metered Building" were derived
from surveys completed by the buildings' electric utilities. The utilities filled out
information on whether the buildings paid a demand charge as a component of their
electric bills. From the 1995 CBECS data, we estimate that approximately 50% of
commercial customers currently have meters capable of measuring demand.
Load control options can be tailored to reduce demand and hence also create a bill
savings. Figure 2.3 displays the number of demand metered buildings as a fraction of the
total number of commercial buildings in the 1995 CBECS survey according to specified
building size ranges. Forty-nine percent of buildings less than 25,000 square feet pay a
demand charge. Seventy-four percent of buildings between 25 and 100,000 square feet
pay a demand charge, and 78% of buildings larger than 100,000 square feet pay a
demand charge, according to EIA data. Figure 2.4 shows the same information with
more detailed building size categories. Forty-three percent (957/2222=43%) of demand-
metered buildings are less than 5,000 square feet in size, and 63% are less than 10,000
square feet. While most large buildings incur demand charges, 22% do not. (EIA 1995)
EIA data indicate that certain categories of building activity dominate the total floor
space of buildings greater than 100,000 square feet: education, health care, mercantile
and service, office, and warehouse and storage.
15
Demand Metered Commercial Buildings as
Fraction of Total, Aggregated Size
100%
90% B Buildings Not
O CD 80% Paying
c (0 70% Demand
o O
f a> 60% Charge
3 N 50%
c/5 40%
tr
c • Buildings
c 30%
o (0 Paying
O CD 20% Demand
c TJ 10% Charge
0)
'5 0%
u
o) m
o. <25 25.001- >100
100
Size, Thousands of Square Feet
Figure 2.3. Demand Metered Commercial Buildings, Aggregated Size Categories.

Numbers in chart reflect total number of buildings, in thousands, in each size category.
For total buildings by size, add categories vertically.
Demand Metered Commercial Buildings as Percent of

Total Buildings, by Size
I B Buildings Not
o * Paying
c (0 Demand
o o Charge
CD
N
a co
• Buildings
o <o Paying
o o» Demand
Z. .E
C
0)
TJ
=
Charge
2 =
0) CO
Q.
Size, Thousands of Square Feet
Figure 2.4. Demand Metered Buildings, Refined Size Categories. Numbers in chart
reflect total number of buildings, in thousands, in each size category.
16
2.2.5 Electric Use in Demand Metered Buildings
According to CBECS Table 10, "Electricity Consumption and Expenditures Intensities,

1995," (EIA 1995) the most electric energy intensive activities per square foot occurred
in buildings engaged in food sales, food service, health care, and office activities. Note
that these building types also have the highest peak demand per square foot, as shown in
Table 2.1. These values were derived from CBECS Table 19, "Distribution of Peak
Watts per Square Foot and Load Factors, 1995."
Load factor is defined as the ratio of average demand or energy use to peak demand.
Ideally utilities would prefer all customers to have a load factor of unity (i.e. constant
diurnal and day-to-day electric consumption). This would mean that electricity-
production capacity would be used most efficiently (because no load-based capacity - as
opposed to reserve capacity - would be idle). Median load factor values shown in Table
2.1 are also derived from CBECS, Table 19 (EIA 1995).
Number Electricity Used Median Peak Median

of per Square Foot Demand (Watts per Load
Buildings (kWh) Square Foot) Factor
Food Sales 254 54.1 14.67 0.463
Food Service 171 36.0 12.67 0.333
Health Care 589 26.5 5.89 0.253
Office 281 18.9 6.00 0.285
Mercantile & Service 117 11.8 4.91 0.249
Education 210 8.4 4.29 0.210
Warehouse/Storage 108 6.4 .222 0.265
Table 2.1. Electricity consumption and demand by building type.
2.2.6 Energy Service Providers and Their Customers
The aggregation market is segmenting into bands similar to those that existed prior to
deregulation. In general, residential buying groups, business associations, and large
commercial or industrial facilities with divisions/plants in multiple locations are the
primary targets of energy marketers.
The focus is on bundling load so as to be able to obtain volume discounts from electricity
suppliers. Some energy services, billing analysis and controls work is being done, but it
does not appear to be the main thrust of suppliers at the moment. Many regulated utilities
have created a separate company from their commercial and industrial customer support
divisions. This enables a utility holding company to participate in both the energy and
the retail markets. The pool of energy service providers includes these marketing
subsidiaries, as well as many new companies. Such companies included, at the time of
the research:
17
• Central & Southwest division: EnerShop
• Allegheny Energy Supply (AES), a division of Allegheny Energy
• New Energy Ventures, which has been purchased by AES
• Sempra Energy Solutions
• Exelon, a division of PECO Energy
• Enron Energy Services
Examples of energy service agreements struck in 1999 include energy management

services for chains of establishments - for example, a hotel, bank, or restaurant chain.
Professional associations are another fertile ground for marketers - both building
managers' associations as well as more specific groups like the New Jersey Coalition of
Automotive Retailers (NJCAR), or the National Frozen Food Association. Educational
and health care facilities offer many opportunities for energy improvement and have been
targets of numerous marketers. Additionally, municipal aggregation is occurring in the
Northeastern United States. Table 2.2 summarizes these findings.
18
Organization Owned By Services Provided Contracts With
Cinergy Affiliate of Design and engineering, energy use Business Associations
Business Cinergy analysis, and energy supply ideas • NH Retail Merchants Association
Solutions
PG&E PG&E Corp. Provide electricity and identify energy Commercial loads with multiple locations
Energy efficiency projects. Will arrange • Marriott hotels and resorts in California
Services financing for upgrades if requested. • Massachusetts High Tech Council Load Aggregation Program
800.982.6887 • Equity Office Properties Trust - 15 Office buildings and parking
facilities in San Francisco, San Diego and Orange County.
• Burger King Franchises in CA —450 restaurants
• Portfolio also includes McDonald's and Carl's Jr. restaurants;
Blockbuster video and music stores; Rite Aid and Sav-on drug
stores; Safeway, Vons and Lucky supermarkets; Neiman Marcus
IBM, Mitsubishi Silicon America and NEC America; Smucker's;
Pepsi-Cola General Bottlers and others.
Exelon PECO Energy Fuel and plant management services; Commercial and Industrial
800.595.1185 on-site energy management and • Wampler Foods
888.393.5661 maintenance; distributed infrastructure • Princeton
planning, construction and • Vision Quest
maintenance; energy efficiency • USX
evaluation and implementation. • Massachusetts HEFA-Health & Educational Facilities Authority (an
Energy procurement and brokering; aggregator of loads and finance provider for energy efficiency
consolidated energy billing and projects)
reporting as well as industrial and
commercial energy supply.
ONSITE Billing analysis, demand side • Business Associations of Food Wholesalers/Distributors (e.g.
SYCOM management, direct access planning National Frozen Food Assn)
and cogeneration services. • New Jersey Coalition of Automotive Retailers (NJC AR)
Wheeled Independent Small Businesses
Electric
Power 50 Charles Several thousand retail customers buying electricity and natural gas.
800.224.9767 Lindbergh Participant in retail electric pilot programs in New Hampshire,
516.390.7600 Blvd. Suite 207 Massachusetts and New York. Most customers are residential or small
Uniondale, NY business consumers.
11553
19
New Energy AES • Promised bill savings to CHW of Large commercial users.
(formerly at least $300k/yr
New Energy • ACWA saved $2.6M between • California Retailers Association
Ventures) 1997-1999 • Catholic Health Care West (San Francisco hospital)
• Shared Savings program with • Association of California Water Agencies
NCGA members 80/20 • Endorsed by Northern California Grocers Association (NCGA)
members/NE
Allegheny Subsidiary of Volume pricing Large commercial and industrial users.

Energy Allegheny • Large Industrial Loads (e.g. NJ Chemical Industry Council)
Supply Energy • New Jersey Chemical Industry Council
888.232.4642 • Pennsylvania State University
• Various PA municipalities
• Brandywine Realty Trust
Sempra Sempra Energy Use WebEncharge,1M an internet- Retail chains with multiple locations across broad geographic areas.
Energy (the merged based product to manage energy • Union Bank of California 252 offices in OR, CA, and WA
Solutions SDG&E and S. information and billing. Offers • Advance Auto Parts - 700 stores in 37 states
California Gas) internet access to energy use and • Penske Truck Leasing - 120 locations in Midwest
analysis information. Information is • City of San Diego, CA - 2400 facilities including Qualcomm
updated daily. Also offer energy Stadium and the San Diego Convention Center
cost savings, in one case, of up to 5
percent, and bill consolidation.
EnerShop Wholly owned • "EnerACT" is a two-way • La Quinta Motor Inns - 19 facilities in Texas and California.
subsidiary of communication system for • Building Owners and Managers Association (BOMA) of Dallas
Central & optimizing energy operations.
Southwest Collects facility information over
Corporation the internet. Information is
analyzed and engineers advise on
operations and energy savings
opportunities.
• Energy aggregation
• Control Technology Services
Table 2.2. Summary of aggregated loads by provider.
20
3. CANDIDATE SITE SELECTION
Our goal for the candidate site selection was to identify instances where (i) load
aggregation had been performed on a multi-building level and (ii) the representatives of
the respective organizations were willing to share detailed information about their
experiences. Through previous contacts and communication within the electric power
sector, we compiled a list of candidates that fit the basic requirements of this project.
Upon review by the PMSC, we formed a short list of case studies with which we
conducted an in-depth review to collect more detailed information. This section of the
report describes the general criteria for identifying the potential candidates and the
information gathered from preliminary site interviews. A breakdown of the detailed
report proposed for Task 1.4 of the project is included in Chapter 4, Case Study
Evaluation.
3.1 General Criteria for Selection

In examining the many sites where electric load aggregation has occurred on the multi-
building level, it was first necessary to establish areas of comparison. Prior to performing
a preliminary interview for a case study candidate, we considered the building type
diversity, duration and level of participation, and the extent of load control employed.
Admittedly, we have found it difficult to assemble information for sites that are remote
(i.e. California). Also, because the aggregation agreement established between an
electric supplier and a customer is often considered proprietary and confidential, many
otherwise suitable candidates were unwilling or unable to reveal the details of their
experience for the purposes outlined in this report.
3.2 Selected Candidate Sites
As required by ASHRAE 1146-RP, six candidate sites were identified as indicated in

Table 3.1. The first three sites were not selected for detailed case study analysis, but are
briefly described in Section 3.3, Preliminary Site Interviews. The remaining three sites
are more fully described in Chapters 5 - 7 .
21
Organization Location Contact Phone
University of Delaware Newark, DE Richard L. Walter, CFM 302.831.2618
Director of Facilities
Management
Wyeth-Ayerst Pearl River, NY Michael A. Greenholtz 914.732.4075
Pharmaceuticals Director of Facility Operations
Fort Dodge Animal Health Catania, Italy Al Forte 973.660.5894
Facility Director of Energy Procurement
Drexel University Philadelphia, PA Bill Taylor 215.895.2827
Senior Plant Engineer
Thomas Jefferson Philadelphia, PA Randolph L. Haines, CEM, CPE 215.503.6099
University Energy Manager
American Home Products Various sites in NJ Wain Stowe 973.660.5894
and PA Director of Energy Procurement
Table 3.1. Candidate sites.
3.3 Preliminary Site Interviews
Following an evaluation of the general criteria and upon receipt of a positive response
from our interest survey, we performed on-site interviews with a representative from each
organization. During these evaluations, we gathered information and separated it into
four distinct sections:
(1) Site Characteristics - includes the name, location and size of the facility, both in
terms of the number of buildings and the overall gross square feet.
(2) Electric Utility Information - occupied the bulk of our conversation, and was the
primary thrust of this phase. This category is left wide open to include any particulars
regarding the organization's usage history, rate structure, competitive negotiations, and of
course, aggregation experiences.
(3) Building Operation - was much less of a concern, as it will be covered extensively in
the next phase. Basic plant information, peak-shaving strategies other than aggregation,
building control and automations systems, and any operational information are outlined.
(4) Contributions to the Project - summarizes those aspects that make each case uniquely
challenging or advantageous.
Please note that the following interview information is included for candidate sites that
were considered but not chosen for the case study evaluation. Information was obtained
in 1999 and 2000 and was current as of that time.
3.3.1 University of Delaware
Site Characteristics
Located in Newark, Delaware, the University of Delaware is the largest private
educational institution in the state. The 2.5 mile-long campus has over 100 individual
22
buildings including residence halls, cafeterias, academic buildings, laboratories,
auditoriums and recreation facilities totaling 6,900,000 gross square feet. Buildings are
grouped in a number of configurations, requiring examination of individual and multi-
building performance to determine the suitability of the site for this project.
Electric Utility Information

With an annual electric energy budget of approximately $7 million, the university is the
largest electric consumer in the area. In fact, the University of Delaware represents one-
third of the entire power supplied by the City of Newark, while the next largest customer
represents less than 5% of the total load. Along with its large relative size, an impressive
load profile and a load factor fluttering around 0.75 gave this candidate a great deal of
bargaining power when it negotiated a conjunctive billing agreement with the city in
1997.
Electric power is supplied at 4160V or 12,470V by city utility poles that are scattered
across campus. Previously, 14 city meters recorded data for the many loops that
distribute power to individual buildings. This arrangement changed in 1997 when the
City of Newark approached the university with a plan for aggregated billing. The city,
which purchased power from both Conectiv and PECO electric utilities, guaranteed to
provide the best possible electric rate and to calculate the demand portion of the electric
utility bill based upon the conjunctive peak of all 14 meters. In exchange, the University
of Delaware agreed to install electronic meters utilizing radio-frequency technology at an
expense of approximately $180,000. Within this contract, it was also agreed that the
responsibility for physical maintenance of the meters and their enclosures would be
shared by the city and an independent maintenance company, while computer and
software expenses were to be absorbed by the university.
The telemetered system replaced the city metering system, but required an initial expense
to the university. Radio-frequency (RF) technology was selected because the long
distances between stations on campus made ground Ethernet wiring prohibitive.
Currently, 19 RF meters (3 are not in service) send information to a central metering
station on campus. From this location, a hardwired Ethernet connection has been
established with the city. The data at this central metering station, available to both
university and city personnel, are used to bill the university. Savings were projected at
15%, which formed the basis of the billing contract for the first year. However, once
actual annual savings were established at 9%, which was lower than expected, the
university paid a disbursement to Newark Power for the difference. Penalties are added
for any individual meter that dips below a 0.90 power factor.
It should also be mentioned that Newark Power purchases its electric energy from
suppliers as a part of an energy utility consortium called DEMEC, which includes
municipal cooperatives for cities in the area but does not include Wilmington. The
University of Delaware represents 20MW of the City of Newark's 60MW load. The
consortium, which has a total of approximately 200MW, purchases its power in bulk
from the regional power pool. In this way, DEMEC also acts as an aggregator, but in the
sense of a power broker with two people and no tangible assets. The result is that the
23
University of Delaware takes advantage not only of its aggregation agreement, but also
the savings afforded to DEMEC as a large player in the regional power pool.
Building Operation
In addition to savings realized as a result of coincident peak billing versus the sum of all
non-coincident peak bills, the University of Delaware has been able to use the data from
its central metering station to take advantage of diversity in demand. Electric energy
savings have been realized by load-control measures including, but not limited to energy
efficient lighting, occupancy sensors, conversion from direct-expansion air handling unit
systems to central chilled-water systems, installation of an energy management and
control system (EMCS) to monitor and adjust set points for individual chillers and air
handling units, and installation of variable-speed motor drives. Through these measures,
the University of Delaware has been able to reduce its annual electric bills by an
additional amount of almost 25%. Though the people and equipment associated with
university growth have caused a slight increase in overall consumption each year, energy
saving measures have helped limit the increase in annual electric utility costs.
Primary equipment is distributed among four central plants, three of which operate with a
building automation system (BAS) to control system performance. One of these plants
uses Andover, two use Staefa, and the last is controlled manually. These four plants
represent the largest of the 16 electric services on campus. Other large services would be
those that include laboratory buildings, which operate at approximately 60% outside air.
Classes year round include two summer sessions attended by only 40% of the student
population, but because many camps and conferences take place in university facilities
during the summer months, the usage does not drop significantly.
Contributions to the Project

As a potential case study candidate, the University of Delaware offers incredible building
diversity. Also, 15-minute demand data are available for all meters for all three years
that the university has operated under the described load aggregation agreement.
Submetering data are available for some of the larger buildings, but only for kWh
consumption. While some effort is required to interpret the individual bills for previous
years, information from the central metering station is readily transferable in electronic
form. Our primary concern is that because only 16 meters service the university, the 6 or
7 buildings serviced by any one meter may have already acted in aggregate. Without
individual building data, it is difficult to construct an accurate picture of how load control
can be applied in the newly aggregated, multi-building environment.
Despite its many diverse buildings, Delaware University was not included in the Case
Study Evaluation because (1) the original metering arrangement was already highly
aggregated, (2) the municipality has changed buildings from one meter to another, and
(3) such a large site would be difficult to document accurately.
24
3.3.2 Wyeth-Ayerst Pharmaceuticals
The 500-acre facility in Pearl River, New York has more than 3,000,000 gross square feet
of building area dedicated to pharmaceutical manufacturing, vaccine production, research
and development. The 2,700 employees on site are scattered among 14 primary
buildings. Despite its large size, this industrial facility offers many unique insights into
the practice of load control and load aggregation. In the past few years, Wyeth-Ayerst
has ferociously attacked the annual energy budget of this and other facilities in a
successful attempt at cutting demand and consumption of electric power.

Since 1992, this candidate site has been able to reduce its total annual electric energy
expenditures by more than 35%. These reductions have been effected by several
calculated steps in what Director of Facility Operations, Michael Greenholtz, refers to as
a "master plan" for energy management. By far the most talked about portion of this plan
is the installation of a 17MW gas-turbine cogeneration facility and the accompanying 15-
year power sales agreement instituted in 1991. Under this agreement, the local electric
utility company purchases all of the power generated by the cogeneration facility,
allowing the turbines to run at maximum capacity. This arrangement provides the site
with significant annual savings and also the steam generation required for many plant
operations. Determined to make the Pearl River Campus as self-sufficient and
technologically advanced as possible, the energy team of Wyeth-Ayerst has implemented
a number of other programs since 1991 to complement their cogeneration capability.
One of the site's largest changes in electric utility direction came in 1995 when Wyeth-
Ayerst joined forces with other local industrial customers to eliminate the Peak Activated
Rate (PAR) tariff of the local utility company. Under this rate structure, the electric
utility could increase usage charges up to $0.50/kWh for 100 peak hours per year. This
resulted in annual penalties of approximately $1 million. Twenty-four hours of notice
were given before the PAR charge was introduced, but often department managers would
not discover the change until only a few hours before the rate took effect. This warning
typically came in the form of a last-minute email message urging individual departments
to limit load wherever possible.
The basis of the complaint was that under such conditions, Wyeth-Ayerst could not
realistically reduce its load to avoid significant penalties. Though extensive measures
were implemented, the net result was that the manufacturing and research facility was
unable to reduce its 27MW load appreciably with the curtailment notice provided. After
much negotiation, the tariff was eliminated, freeing a significant portion of the
company's energy budget for use on plant improvement projects, which are discussed
under the building operation section that follows.
Another major step was the purchase of the primary substation that serves the campus.
Upon acquisition of the transformer substation, Wyeth-Ayerst was billed according to the
25
much lower high-voltage rates, but still held accountable for distribution charges. In
addition to these savings, the organization has been negotiating electric power costs with
generation facilities since 1997. While New York has not yet finalized deregulation
legislation, Wyeth-Ayerst has been taking advantage of its size and bargaining power by
reducing its usage rates prior to unbundling of the electric utility statewide. Also, as a
member of the Industrial Energy Users Association (TEUA), Wyeth-Ayerst has
collectively bargained for an even better usage rate. As part of this 50MW block, the site
has interviewed potential Energy Service Providers (ESP's) and negotiated a competitive
aggregation agreement that is signed individually by each member of the coalition.
Under this arrangement, the facility is guaranteed a stable rate so long as they stay below
the threshold stated in the agreement. For overuse, a penalty is assessed in the form of
charges from the local utility, Orange and Rockland Utilities Inc., at the standard market
rate. While the energy manager of the site admits that Wyeth-Ayerst is probably not
getting the best electric power rate possible, he notes that it involves minimal risk and a
measure for aggressive targets to be set in the future. By adjusting annual energy budgets
and aggressively seeking alternatives to electric utility billing rate structures, Wyeth-
Ayerst is able to ensure that its annual electric energy costs are as low as possible.
Building Operation
Because many of the pharmaceutical processes and animal facilities on the campus are
restricted by safety guidelines of government agencies including the FDA and ALAC,
load control measures were typically restricted to shutting off non-essential lighting
systems and using backup diesel generators. Following the NOx/RACT (nitrous
oxide/reasonable available control technology) regulation, even the option of using these
backup generators was eliminated. In response to the PAR charge, Wyeth-Ayerst was
forced to try many other creative strategies, including precooling of buildings, treating
wastewater with liquid oxygen instead of a blower, and sacrificing occupant comfort.
While these load control devices did not give Wyeth-Ayerst the ability to successfully
curb demand upon request, they did give plant managers excellent exposure to the many
ways that buildings can be optimized, individually and in aggregate.
As part of the learning experience from this and other energy battles, Wyeth-Ayerst has
incorporated load control measures and optimization of electric power usage into the
company infrastructure. As a matter of good practice, company representatives evaluate
the options for new energy projects and select those that offer the most substantial
savings in the long term. Plans for the next phase of this energy conscious initiative
involve an online monitoring system for electric energy costs and building performance.
Armed with computerized process monitoring, cooling tower operation, boiler blow
down control, hot and cold deck temperature information, and real-time metering that
breaks total actual electric usage into categories, plant managers will be able to monitor
their systems on an ongoing basis and zoom in on the areas where they can save the most
dollars. These metering and automation improvements will also give Wyeth-Ayerst the
ability to show future ESP's their actual profile, with the understanding that customers
who can communicate in such real terms will be the ones who are offered the most
attractive rate structures in the competitive wake of deregulation.
26
While the existence of a cogeneration facility may pose some hurdles for analyzing
electric utility information and building dynamics on the campus, the benefits of this site
cannot be overlooked. Wyeth-Ayerst's zealous action in demand-limiting strategy as
well as plans for future improvements are clearly in support of the objectives of this
project. As part of their long-term master plan, the energy management team would like
to move to an online format, which combines an EMCS with real-time load-monitoring
capability.
Wyeth-Ayerst Pharmaceuticals was not chosen for the Case Study Evaluation because (1)
individual building use was difficult to discern, (2) meter information prior to
aggregation was unavailable, and (3) the primary reason for entering an aggregation
agreement was not to control multi-building load, but to create a coalition to battle unfair
legislation and secure lower electric pricing.
3.3.3 Fort Dodge Animal Research Facility
In talking with representatives from American Home Products' Corporate Headquarters
in Madison, New Jersey, we were informed of this international subsidiary in Catania,
Italy. While specific information regarding the gross square footage of the site is lacking,
it is known that the facility is used for bulk production, toxic tablet manufacturing, and
blending operations. Our contacts have volunteered to supply additional information in
subsequent interviews.

The key element of this site is the complicated rate structure that dominates electric
power markets in Europe. In the case of Catania, the current real-time pricing structure is
broken into categories from Fl to F4, each of which has multiple rates associated with it.
These combinations of available rates vary according to the realized supply cost of
electric power. Ninety meters, representing every major electric user in the area, are tied
to a central computer. From connections with this Building Management System (BMS),
each site can control its own cost of goods by managing its load in response to actual
market dynamics. With this complex tariff in effect, the incentive for demand-side load
management is great. Annual electric costs for the Catania facility were approximated at
$2 million, but no specific information was given regarding usage, demand or realized
savings. However, we do know that local users generally budget electric power as 15%
of their total cost of goods, making it a large concern for commerce and industry.
Building Operation
Since 1995, the Catania facility has implemented a variety of energy savings projects,
many of which directly affect their electric usage. The first and one of the most
expensive was the installation of the automatic supervision for energy consumption
discussed under electric utility information. Other building improvements include an
optimization of the cooling-tower-water-distribution system, installation of variable-
27
speed motors in cooling tower fans, and extended BAS, installation of air-fuel
modulation system in hot water boilers, an HVAC steam and condensate-recovery
system, installation of heat-recovery systems in air-conditioning system, improved hot-
and chilled-water distribution, and a list of more recent projects that will be made
available in any subsequent interviews.
It should be noted that under the automatic-supervision project, a significant amount of

monitoring and control equipment was added to the facility. This phase also included the
period when the site negotiated with the electric utilities and optimized its power-supply
contract. Upon completion of this project, energy managers began issuing daily, weekly
and monthly reports that enable them to get a very clear picture of energy usage. From
this clear picture of the site's profile, individual departments can make decisions for how
operation should be changed in order to minimize electric costs.
Presently, a proposal is being entertained regarding the construction of a 2.6MW

cogeneration plant for the site. The plan is schedule to take effect some time in the year
2000.

The highly developed electric-utility market and the sophisticated rate structures in Italy
make the Fort Dodge facility a valuable resource for information regarding load
aggregation and plant optimization. By examining the progression from bundled service
to this competitive electric power market, we are confident that we can form a clearer
picture of what future practices can be expected for building operation in an aggregated
load.
While insight was gained on the development of real-time pricing and advanced rate
structures, the Fort Dodge Animal Health Facility was not included in the Case Study
Evaluation, primarily because of its remote location.
28
Organization Primary Total Gross Peak Total Annual
Buildings Square Demand Electric
Footage Budget
University of Delaware 100 6,900,000 21MW $7,000,000
Wyeth-Ayerst Pharmaceuticals 14 3,000,000 25MW $7,700,000
Fort Dodge Animal Health — — — $2,000,000
Facility
Drexel University 30 2,800,000 6MW $3,000,000
Thomas Jefferson University 16 4,000,000 21MW $5,500,000
American Home Products — — — —
Table 3.2 Summary of Candidate Sites. All data reflect approximate totals for the year
1999 from preliminary candidate interviews.
29
4. CASE STUDY EVALUATION
Our goal for the case study evaluation was to expand upon the information obtained
during the preliminary interviews by carefully documenting the aggregations of the
following three sites.
Organization Location Contact Phone

Drexel University Philadelphia, PA Bill Taylor 215.895.2827
Senior Plant Engineer
Thomas Jefferson Philadelphia, PA Randolph L. Haines, CEM, 215.503.6099
University CPE
Energy Manager
American Home Various sites in NJ Wain Stowe 973.660.5894
Products and PA Director of Energy
Procurement
Table 4.1 Case study sites.
Following approval by the PMSC in February 2000, we conducted an in-depth review for
each case to collect detailed information on the control of multiple buildings in an
aggregated load. While our approach to each site was the same, the information that we
obtained was slightly different. For this reason, we have chosen to document the first
two case studies using a similar format and to deviate from this format to present the last
case study. The following sections will describe the goals of the case study evaluation,
our approach to documenting the sites, and a summary of the presentation format for both
Drexel and Thomas Jefferson Universities.
4.1 Objectives
As outlined by our original proposal, the primary objective of the Case Study Evaluation
was to identify situations and conditions under which aggregating individual building
loads is attractive for managing total, multi-building load. Secondary objectives
therefore included (1) documenting site characteristics such as building size, location and
type of business, (2) obtaining load profiles, utility data, metering guidelines, and rate
structures, (3) outlining the primary loads for each metered group as well as strategies for
shifting, curtailing or eliminating loads, and (4) determining the financial implications of
aggregation. By breaking the information down into these four categories and presenting
it in a clear and readily comparable format, we have created a platform from which we
may step into Phase 2 of the project. Our presentation format, as discussed in Section
4.3, is designed to make the case study evaluations as straightforward as possible.
30
4.2 Documentation Methodology
Ultimately, we are interested in how two or more buildings may be grouped together to
take advantage of their respective differences in profile. It is not enough, however, to
simply know what the monthly, hourly or quarter-hourly electric load is in order to
predict how best to control multi-building load. Rather, we need to have detailed
information regarding end-use loads, preferably for each of the buildings involved, so
that we may target those areas where the most significant savings may be realized.
Unfortunately, few buildings currently have the systems in place to accurately monitor
end-use load for a given building. Since this information was unavailable for any of our
case study sites, we decided to identify and present the building use type, broken down
into percentages, for all the different areas in a building. By doing so, we enable the user,
whether it be an aggregator or building owner, to identify what type of building is being
considered - classroom, laboratory, or possibly a mixture of several others - and then
compare it to the information we have gathered.
If the square footage and building use type are known, it may then be possible to generate
a generic load shape, similar to a known building. By comparing this model with
recorded utility data, even if only a monthly kW and kWh value, the user should be able
to correct the annual electric profile curve sufficiently so that when two or more of these
virtual buildings are combined, the utility bills calculated for their combined profile will
approximate the savings of a true aggregation.
4.3 Presentation Format
To simplify the presentation of information gathered from Drexel University and Thomas
Jefferson University, a summary sheet has been put together for each metered group. The
following sections describe how these summaries were assembled and clarify some of the
terminology used to present the data. The last page of each Metered Group Summary is
reflective of the aggregated group of buildings. In this section, a Building Summary,
rather than a Metered Group Summary, is shown. The primary and secondary uses are
not given because they have already been represented by the preceding information
sheets. Also, rather than referring to the Building Locator Map, it would be advisable to
refer to the campus map which precedes Section 5.3 and Section 6.3.
The summary sheets are broken down into five basic sections, each of which is called out
in Figure 4.1. A brief description of each section is given following the figure.
31
•mater- •*""
p--^r-sr
: JUBDIOBOIIJ frvt (ra>
L
0
r * " « f t » «n.uil £.»<J ™ o "
: plat H T O MJt •ftinltaa) & t
R
i M.Otrf I t W * b C . 0 0 0 * 1
(1) Metered Group Details

i"
(2) Building Locator Map
*BSE=:
(3) Building Use Summary 1 an BKHJUI (Muy aba bat
tadif, cadlai • h»lo*«i4 m l
! 11 H o * I » I U bare n p a n * «•
i—i 4 7]
(4) Metered Group Use Breakdown
1
I isTctaqi lopmrcbatlirfnn)
1 (•d^waailr.BuibatcaBtBaoi
s*
(5) Chart of Metered Group Use

BuiU^f « ] :S«0»l4« IJ
•v. tfipdV
;WM-T
iCImaaai
s S W H H « UBMIT
idnW / l
tWî-Hi^t —
4 1 .ft™-™ ,Z
bWlwwtWI* ^** i? «Clamaaai
- J
•*•*" CflSee
UtNDOi.HM • met Ldwiami Clnin»
T+n 1 C-mi lû« r»F***>r SS33K
«1t Office
2*' b»i ISJkCtt

lltt R a » . « A
.iftjtitt
«3u4
4V. CEUNBUB
*14 dtbo MJK»
! j
Figure 4.1. Metered group summary key.
4.3.1 Metered Group Details
This short list of figures gives specific details about the entire metered group. Details
include the number of buildings served by the meter, the combined square footage of all
the buildings, equipment information, utility data, and any comments about the group.
The equipment information is broken down into primary heating and primary cooling.
This section includes any heating or cooling plants and the total chiller tonnage, if
known. Utility data include annual demand peak, energy demand index, annual electric
usage, and energy use index. The terms are described below:
Annual Demand Peak

Rate structures are typically configured to bill customers according to an increased rate
for a single high point during the year. In addition, they tend to inflate the rate based
upon a seasonal peak, usually during the summer months, for an entire year. For this
reason, we represent the demand of each group in the form of the Annual Demand Peak.
This number represents the highest single monthly kW peak for the block. Note that this
peak represents the highest single demand for a given meter, and therefore a specific
metered group of buildings, for the entire year.
Energy Demand Index

By dividing the Annual Demand Peak by the total gross square footage for the entire
metered group, we define an area-normalized measure called the Energy Demand Index
(EDI). This normalized value shows the W/ft2 for the period of highest demand of the
metered group. This is not representative of the actual demand for any specific area
32
within the group. In fact, it may vary drastically from building to building and is not an
indication of building optimization or overall performance. It is, however, a useful term
for comparing the relative maximum demand for two metered groups of different areas.
Annual Electric Usage

Quite simply, the Annual Electric Usage is the total kWh used by the metered group for
an entire year. This information can easily be obtained from utility bills or on-site kWh
meters.
Energy Use Index

Similar to the Energy Demand Index, the Energy Use Index is an area-normalized value
that divides the Annual Electric Usage by the total gross square footage to give an
indication of electric use (in kWh) per square foot per year for the entire metered group.
The EUI is a much better indicator of the actual use of a building or metered group and is
widely used to compare buildings of differing size and use type.
4.3.2 Building Locator Map
An entire metered group involves a group of buildings that all receive their power from
one substation or one meter. The locator map is used to show how these buildings relate
to each other, their footprints, and to link information in the Building Summary below
each map.
4.3.3 Building Summary
Just as all the metered groups are combined into one aggregated load to characterize their
overall performance, so must each group be broken down in order to consider its
individual components. In the Building Summary, we list each building by number and
name and give information regarding its total gross square footage, the primary use type
and secondary use type. Where actual square footage measurements were unavailable,
approximations were made from building floor plans. The use types are chosen from a
list of thirteen designations, including classroom, office, and circulation spaces. Use
types are based on the U.S. Department of Energy's "End-Use Load and Consumer
Assessment Program" (ELCAP) report on commercial building electricity end uses.
Categories have been added to account for the diverse building types included in the
'University' building category. 'Mechanical', 'Circulation', and 'Other' have been
included as categories of use because they contribute to the programming of total gross
square footage. A better approximation might be to consider net assignable square
footage (NASF) and eliminate these categories as unconditioned spaces, but this would
require a much more detailed survey of each building. The list of thirteen use types may
also need to be revised to accommodate additional building types or to combine use types
that have similar end-use profiles.
33
4.3.4 Metered Group Use Breakdown
With a square footage breakdown for each use type for each building, it was easy to sum
each category to find the total square footage for the entire metered group that was
designated for each use. The Metered Group Use Breakdown is simply a table showing
the square footage and corresponding percentage for each use type within the metered
group. Note that the sum of all these values should be identical to the total gross square
footage and the sum of the individual building areas.
4.3.5 Chart of Metered Group Use
This pie chart contains exactly the same information as the Metered Group Use
Breakdown, but it distinguishes between use types much more clearly. For each metered
group, a particular use type will have a similar hatch. That is to say that if the piece of
pie showing 'classroom' use is a diagonal hatch for Metered Group #1, the classroom
space for Metered Group #2 will also be indicated by a diagonal hatch. Note that all use
types may not be present for each metered group.
In the case of a metered group with only one building, the chart will show the exact
breakdown of all the spaces within that building. However, where two or more buildings
are grouped together, their use types may become more difficult to discern. For example,
a building that is primarily 'Classroom' space combined with a building that is primarily
'Laboratory' space may result in a pie chart that shows equal portions for each of these
categories. Similarly, because it is present in all building types, the 'Mechanical' use
type may become a very large chunk of the groups use, possibly even the largest portion.
In any case, the chart allows a quick assessment of the diversity inherent in the grouping
of buildings, and also indicates those areas that may contribute heavily to the electric
load.
34
5. DREXEL UNIVERSITY
Drexel University has a wide variety of building types available for research. Even more
importantly, buildings of like use are often in close proximity to one another. For this
reason, even in cases where individual building information is not present, a use profile
may be relatively easy to discern, especially in those areas where a complex of similar
buildings is served by one mechanical plant. Also, the Senior Plant Engineer has
extensive knowledge of the optimization techniques that have been used to shave demand
peaks in the past eight years. The BAS data for the primary mechanical plants on campus
are readily available in electronic format. Drexel has been most helpful in sharing
information regarding load control strategies, utility rate information and personal
experiences with load aggregation and deregulation.
5.1 Site Characteristics
Located in the University City District of Philadelphia, Drexel's urban campus meanders
from 31 st Street to 34th Street along both the north and south sides of Market Street.
Though many buildings create the grassy islands that float about the site, only a portion
of these are considered in our theoretical aggregated load for the following sections.
These 23 buildings represent just over 2,190,000 square feet and include dormitories,
offices, laboratories, classrooms, two theaters, a parking garage, a library, and an athletic
center. Restricted to a few city blocks, the residence halls are grouped on one end of
campus while research and laboratory facilities are similarly grouped on the other end of
campus. As part of the University City community, Drexel is connected to the Trigen
steam loop along with the neighboring University of Pennsylvania. Figure 5.1 shows the
layout of the campus, as well as the location of the individual buildings. For a
description of each building, refer to Section 5.3, which shows information for each of
the metered groups on campus.
35
De«3 UN\esity
GrrpLBlVtp
Figure 5.1. Drexel University campus map.
36
5.2 Rate Structure, Utility Profile and Metering
Drexel University is served by nine substations, each with its own meter. For this reason,
we elected to break the campus up into nine discrete zones, or metered groups, as shown
on the map in Figure 5.1. We approximated the peak demand as though individual
meters had been combined into an aggregated load. For the purposes of this report, we
extracted monthly usage and demand information for each of the nine meters to be
considered as part of the entire site load. The buildings served by each meter and the
associated total gross square footages are given in Table 5.1 below. The normalized
value of W/ft shown in this table was calculated as a rough estimate for comparing the
relative demand of the metered groups.
Buildings Total Gross Mean Monthly

Meter # Served Square Footage kW Peak W/ft2
1 5-6-8-9-11-12 533,840 1,685.58 3.157
2 1-2-3-4-7-27 483,343 1,734.10 3.588
3 13-14 221,036 661.62 2.993
4 15-16 337,600 470.62 1.394
5 55 128,355 241.67 1.883
6 10 44,860 161.83 3.608
7 23 50,630 306.83 6.060
8 20-22 210,200 1,010.05 4.805
9 19-21 185,500 354.42 1.911
2,195,364 6,626.72 3.019
Table 5.1 Drexel University metering groups.
Two of these meters, #1 and #2, are dual services, which means that physically there are
two meters present, but the sum of their readings is billed as though it came from a single
meter. Each meter of a dual service is sized for half of the entire zone's load, and both
are supplied with power from separate local distribution centers. The practical
justification for this design is that if a power outage should occur at a distribution center,
the entire zone may be switched to the other meter and primary building functions
maintained. The likelihood of two independent distribution centers experiencing outages
simultaneously is remote. For this reason, a dual-meter service functions much the same
as an individual meter, and for the purposes of this report, will be treated as such.
These nine electric services distribute power to the clusters of buildings across campus,
but individual buildings are only sub-metered for usage (kWh), not for demand (kW).
The result is that each of the zones on campus is processed as a single unit. The bills are
processed without regard to individual building usage or other metered groups. Since
1996, Drexel has participated in a long-term service billing agreement, but the specifics
of this agreement are confidential. Therefore, we have projected information from
37
Drexel's historic electric rate structure1 and a theoretical aggregated load. Unlike the
historic rate structure, the theoretical aggregated load assumes that all nine zones are
considered simultaneously, in aggregate, for the entire month. This block is then
processed according to the same rate structures as the individual buildings, but for one
monthly kW and kWh value.
The result is an aggregate billing arrangement that takes advantage of the fact that each
zone peaks at a slightly different time of day. Our earlier projections indicated that by
entering into such an aggregate billing agreement, Drexel University could save
approximately 8% on its monthly electric energy bills. The final projection for savings is
clearly presented by comparison of historic billing information with current kW and kWh
readings in Section 5.5.
For each metered group of buildings, we used individual monthly peak kW demand
values to arrive at a summed kW peak demand four our proposed building group. This
was compared to the single conjunctive peak that would occur if all buildings were
pooled together. Figure 5.2 illustrates how the sum of the peaks and our calculated
aggregate demand fluctuate over the course of a one-year period.
Figure 5.2. Drexel University theoretical peak demand of aggregated load vs. sum of
individual peaks.
To form a more realistic picture of Drexel's overall profile, the monthly kW peak (Table
5.1) and monthly kWh (Table 5.2) readings were used to compute the load factor for
every metered group and every month. In addition, we calculated a load factor for the
theoretical aggregation, indicated by the coincident values. Table 5.3 shows the monthly
values for load factor on Drexel's Campus. Finally, these load factors are plotted in
Figure 5.3 and Figure 5.4. Shaded sections in tables are areas where no data were
available.
1
As of 1996, PECO Energy provided electric power to Drexel University under Rate HT, the commercial
high-tension rate.
38
Individual Monthly Usage (kWh) Monthly
Meter # Jul-99 AUO-99 Sep-99 Oct-99 Nov-99 Dec-99 Jan-00 Feb-00 Mar-00 Apr-00 Mcy-00 Jun-00 Mean (kWM
1 1,006,705 955,031 961,571 778,195 696,372 632,614 768,158 722,033 811,730 695,359 787,112 944,000 813,240.00
2 1,144,709 1,074,424 1,106,609 830,983 718,987 788,053 908,595 814,007 816,826 643,837 730,569 1,130,000 892,299.92
3 370,000 361,000 390,000 331,000 294,000 299,000 311,000 292,000 332,000 272,000 306,000 317,000 322,916.67
4 342,427 328,494 329,119 275,449 216,454 220,802 250,656 199,642 244,690 192,858 236,159 332,500 264,104.17
5 179,900 159,500 167,200 113,100 87,400 91,400 95,600 100,400 98,100 80,900 108,700 145,700 118,991.67
6 116,400 112,400 127,900 0 131,100 30,600 106,400 64,500 70,400 61,000 90,100 79,300 82.508.33
7 102,600 98,300 98,100 92,300 -9,700 81,200 190,400 108,200 112,300 77,100 91,100 93,400 94,608.33
8 301,400 301,400 267,600 326,800 339,800 381,700 511,900 736,500 512,600 332,100 382,900 378,000 397,725.00
9 111,792 113,741 138.437 176.298 143.637 150,531 178,691 234,390 189.664 141,822 199,635 62,900 153.461.50
Sum 3,675,933 3,504,290 3,586.536 2.924.125 2.618.050 2.675,900 3.321.400 3,271,672 3.188.310 2,496.976 2,932,275 3.482.800 3.139,856
Table 5.2. Drexel University monthly electric use.
Individual Monthly Load Factor Monthly

Meter # Jul-99 Auq-99 Sep-99 Oct-99 Nov-99 Dec-99 Jon-00 Feb-00 Mar-00 Apr-00 May-00 Jun-00 MeanLF
1 0.68 0.67 0.74 0.59 0.57 0.57 0.72 0.76 0.76 0.68 0.54 0.69 0.66
2 0.74 0.71 0.81 0.61 0.63 0.68 0.76 0.78 0.74 0.69 0.51 0.79 0.70
3 0.74 0.72 0.81 0.61 0.58 0.64 0.67 0.67 0.72 0.62 0.58 0.66 0.67
4 0.88 0.82 0.89 0.72 0.63 0.73 0.84 0.75 0.84 0.69 0.55 0.89 0.77
5 0.52 0.58 0.68 0.66 0.64 0.37 0.57
6 0.30 1.04 0.69 0.66 0.58 0.45 0.62
7 0.40 0.39 0.45 0.35 0.74 0.74 0.43 0.39 0.65 0.30 0.36 0.47
8 0.57 0.64 0.72 0.59 0.49 0.42 0.41 0.69 0.56 0.58 0.50 0.52 0.56
9 0.68 0.72 0.75 0.61 0.68 0.52 0.52 0.77 0.74 0.69 0.54 0.20 0.62
Coincident 0.81 0.79 0.88 0.65 0.68 0.64 0.71 0.79 0.74 0.69 0.54 0.73 0.72
Table 5.3. Drexel University monthly load factor.
40
Figure 5.3. Drexel University Monthly load factor for metered groups 1-5.
Aug-99 Sep-99 Dec-99 Jan-00 Apr-00 May-00

Month
Figure 5.4. Drexel University Monthly load factor for metered groups 6-9 and
aggregated load.
In 1998, following electric utility deregulation in Pennsylvania, Drexel University sought

to further reduce overall energy costs. The university experienced an average monthly
load factor as high as 0.70 and had a favorable electric usage profile on a diurnal and
seasonal basis. Historically, this was because the north end of campus, primarily
residence halls, dominated the university demand in winter while the south end of
campus peaked during the summer months. These characteristics made Drexel a target
for load aggregators who offered a decreased rate for power generation and supply to
university officials if they would sign a supply contract for up to three years. Drexel
declined, assuming correctly that it could shop for a better rate on its own than the ESP's
could offer as part of a larger portfolio.
41
In the end, Drexel University found it better to negotiate an agreement with its local
distributor and an independent supplier than to allow and ESP to act as the aggregator.
With its rate structure and utility information fixed, Drexel then focused on how it might
change building operations and dynamics to take full advantage of the opportunities
presented by aggregation.
42
5.3 Metered Group Summary
Metered Group #1 - Drexel University
Number of
buildings: ITO
Total square 533,840 '///////A'-
footage: ''/////'////A 11
Primary heating: Steam from university loop and yy/y/y//y/x
/ y y s S f s S s j
supplemental HW heat.
Primary cooling: 2 x 450 ton central chilled water plant
serves 70% of buildings 5, 6, 11, and 12
with the remainder from DX units.
Single 300 ton chilled water loop serves
buildings 8 and 9. //////////4
Annual demand 1,986 kW 'V////////77L
peak:
y/777>\
Y//////////////////A
• • • ' - • ' • • ' / / / / / ,
Energy Demand 3.72 W/sq. ft.

Index: Y/V/A//AS/>y//
///////// >'' ^- /
Annual electric 9,758,880 kWh Y/////A ///
usage:
Energy Use Index: 18.28 kWh/sq. ft. / T
Comments: This metered group also has backup iZA ^k///y////' A
chillers in buildings 9 and 11. These
units have capacity to serve their
respective buildings independently, but
n
have not been used in several years.
No. Building GSF Primary Use Secondary

Use
5 Stratton Hall 61,060 Classroom Laboratory
6 Korman Center 69,400 Office Classroom
8 Creese Student Center 65,490 Restaurant Office
9 MacAlister/Mandell Theater 152,830 Office Theater
11 Matheson Hall 74,310 Classroom Office
12 Disque Hall 110,750 Laboratory Classroom
26% Office 142,000

21% Classroom 111,000
3% Retail 15,000
11% Restaurant 60,000
14% Laboratory 77,000
6% Theater 30,000
8% Mechanical 42,500
4% Circulation 19,500
7% Other 35,850
533,840
43
Number of
buildings:
Total square 483,343
footage:
Primary heating: Steam from university loop and
supplemental HW heat.
Primary cooling: 2 x 200 ton chilled water plant serves
buildings 1 through 4 for y,
approximately 55%, 2 x 300 ton
Annual demand
peak:
chilled water plant serves 28% with
the remainder being DX units.
2,083 kW ^Â^y7^^^^22z2 •I
Index:
Annual electric 10,707,599 kWh
usage:
Energy Use Index: 22.15 kWh/sq. ft.
Comments: The unusually high cooling load for
this group is attributed to the animal
research facility in the basement of
building 27. Additional laboratory and
data center spaces in buildings 7 and
27 also explain the skewed chiller
tonnage distribution.

Vse
1 Main Building 178,360 Office Warehouse
2 Randell Hall 48,680 Classroom Office
3 Curtis Hall 61,600 Classroom Laboratory
4 Alumni Laboratory 45,630 Laboratory Classroom
7 Commonwealth Hall 87,370 Classroom Office
27 Lebow Engineering Center 61,703 Laboratory Mechanical
Total Gross Square Footage 483,343
35% Office 165,000

6% Warehouse 30,000
5% Data Center 25,000 35%
5% Other 25,843
483,343
11%
24%
44
Number of
buildings:
footage: '^——^—m^
Primary heating: Steam from university //////////A
loop.
s/Z/A
•UÛAi^^4Ai^ '////A
Primary cooling: 200-ton chiller serves
building 13 while building
14 has 40 tons of DX
cooling only.
Annual demand 725 kW
peak:
Index:
usage:
Energy Use 17.53 kWh/sq. ft.
Index:
Comments: It should be noted that the
120,000 square feet listed
for theater includes the
gymnasium, which
functions as a large
auditorium space.

Use
13 NesbittHall 82,000 Classroom Theater
14 Physical Education Center 139,036 Theater Office
11% Office 25,000

55% Theater 120,000
1% Data Center 2,500
5% Circulation 12,000 16%
4% Other 8,036
221,036
55%
45
Number of
buildings: s r ' < • t J t , f 'T\' ' - ' ' J

footage:
Primary heating: Steam from university loop. 16 15
Primary cooling: Building 15 is cooled by a 2 x 160 ton 77L Y7A
chilled water plant, with one chiller «« /- y / /•* /- /•
running during the day and two at

night to provide precooling. 2 x 20 ton
chilled water plant is used to condition
4
the first floor of building 16.
peak:
Index:
usage:
Comments: As anticipated, because this metered
group is dominated by the large garage
space in building 16, the Energy
Demand Index and the Energy Use
Index are relatively low.
No. Building GSF Primary Use Secondary Use

15 W.W. Hagerty Library 100,000 Library Office
16 General Services/Parking 237,600 Warehouse Office
10% Office 35,000

3% 2% 10%
3% Classroom 10,000
63% Warehouse 210,000
13% Library 45,000
2% Mechanical 7,500 13%
2% Other 5,100
337,600
63%
46
Number of
buildings :
Total square footage: 128,355
Primary heating: HW heat through steam conversion.
Primary cooling: DX air conditioners serve entire building with 2
cooling towers on roof.
peak:
Index:
usage:
Comments: The restaurant on the sixth floor of this building
is run just like a commercial kitchen making this
a unique component in the otherwise classroom
dominated building.

55 Academic Building and Annex 128,355 Classroom Restaurant
14% Office 18,000

6% Library 45,000
9% Warehouse 12,000
3% Laboratory 4,000
6% Mechanical 8,000
Other 355
128,355
34%
19%
47
Number of
buildings: W////Z,
Af ' ' '
Primary heating: Steam from university loop.
Primary cooling: 118-ton electric chiller.
peak:
Index:
Annual electric 990,100 kWh
usage:
Comments: Despite its small size, this building has a
significant electric energy load, probably due
in large part to the electric chiller, which
provides cooling for the entire building.

10 Rush Building 44,860 Classroom Office
22% Office 10,000 3%

54% Classroom 24,000 w
4% Mechanical 2,000
3% Other 1,360 11%
44,860
54%
48
Number of
buildings:
Total square 50,630 7 f//// 7/777'
footage:
Primary heating: Gas heat.
Primary cooling: Rooftop units with DX cooling.
peak:
Index:
usage:
Comments: Hess is a research laboratory and structural
testing facility. The Energy Demand Index
is relatively high, but we expect this is
because of the density of equipment within
the building. A large portion of the square
footage remains ambiguous, either because
we were unable to determine the use, or
because it did not fall into one of our
predefined categories.

23 Hess Laboratories 50,630 Laboratory Office
20% Office 10,000 12%

10% Warehouse 5,000
5% Mechanical 2,500
12% Other 6,130
50,630
10%
49%
49
Number of
buildings:
footage:
Primary heating: Building 19 included on the university
steam loop, but building 21 is an electric
dormitory. Individual fan-coil units
provide heat through electric resistance.
Gas boiler used for domestic hot water.
Primary cooling: 250-ton centrifugal chiller for building 19.
Building 21 has DX coils in fan-coil units.
Annual demand 1,695 kW
peak:
Index:
usage:
Comments: Built as a temporary structure, building 21
minimized initial costs by installing
electric fan-coil units for heating and
cooling. The result is quite evident in the
unusually high demand and use figures for
this group. Student kitchens and lounges
make up the 10% block of restaurant for
the group.

19 Kelly Hall 89,500 Housing Restaurant
21 Myers Hall 96,000 Housing Restaurant
10% Restaurant 19,000 2% 4 % 1%

10%
83% Housing 155,000
2% Mechanical 3,000
1% Other 1,000
185,500
83%
50
Number of
buildings:
footage:
Primary heating: Electric resistance heating coil in fan-
coil units. Building 20 has gas boiler -77V/////////M
. y y if„„.z...zBBI,
for domestic hot water.

Primary cooling: 96-ton chiller for core of building 22,
remainder is cooled by DX, through-
the-wall units.
peak:
Index:
usage:
Comments: Similar to Group #8, these two
dormitories are dominated by lodging
space, but also have kitchens and
lounges for students.

20 Calhoun Hall 85,200 Housing Restaurant
22 The Towers 125,000 Housing Restaurant

79% Housing 165,000
3% Mechanical 7,000
1% Other 2,700
210,200
79%
51
Aggregated Group - Drexel University
Number of 23
buildings:
Total square 2,195,364
footage:
Primary heating: Steam from university loop supplemented
by HW conversion and electric fan-coil
units.
Primary cooling: Lower campus served primarily by
individual chiller plants. Upper campus
dominated by DX units in dormitories.
peak:
Index:
usage:
Energy Use 17.16 kWh/sq. ft.
Index:
Comments: When compared with other case studies, De«3 Uh\/esity
Drexel's index numbers are favorable.
The Energy Demand Index and the annual GjTJOLBlVtp
mean monthly load factor are indicators of
an excellent electric consumer, making
Drexel a target for load aggregators.
Group GSF
1 533,840
2 483,343
3 221,036
4 337,600
5 128,355
6 44,860
7 50,630
8 210,200
9 185,500
Total Gross Square Footage 2,195,364
17% Office 405,000

1% Retail 15,000
12% Warehouse 253,000
15% Housing 320,000
2% Library 45,000
7% Theater 150,000 15%
4% Other 87,364
2,195,364
12%
52
5.4 Building Operations and Dynamics
Various load control strategies have also been applied to further augment the savings
from the long-term billing agreement. This resulted in an additional 10% reduction in the
total annual electric energy cost. These measures include HVAC equipment replacement,
curtailing of non-priority loads, changing of operation set points for chiller and AHU
operation, alternate fuel selection, an energy efficient lighting program, and even pre-
cooling in the several university buildings. One demand-control strategy is the active
pre-cooling in the W.W. Hagerty Library (Building 15). The mass of the building and
library stacks is cooled to below comfort levels so that as occupants and equipment warm
the building into the afternoon, less cooling capacity is required to maintain design
temperatures. Other strategies include online monitoring and control of equipment set
points. As the building approaches preset demand targets, room temperatures are raised
and motor activity reduced to limit peak demand.
Drexel has also actively limited demand by linking BAS systems for multiple buildings
on the same electric meter. These systems pass information back and forth, and then
adjust the cooling plant to serve each building independently. During peak summer
conditions, this enables cooling to be reduced in buildings that are relatively comfortable
while increasing cooling in hotter buildings, all while staying under an established
electric demand ceiling.
Utilizing the BAS systems in place in most of the campus mechanical plants, the majority
of optimization has been carried out by the Senior Plant Engineer. This manual
manipulation has been carried out in the form of unique software routines that are fed to
the Andover control systems for several of the buildings as well as adjustments based on
professional experience. The emphasis has clearly been to reduce overall energy costs,
both for demand charges and consumption. These comprehensive building operation
adjustments may provide a very useful tool in the next phases of this project.
This is a prime example of a peak shaving strategy employed under our theoretical
aggregated load that would not be in effect if the buildings were considered individually.
The buildings are each operated so that they have a minimal impact on the total energy
bill, and monitoring and controlling multiple buildings further optimizes the aggregated
group.
5.5 Financial Implications of Aggregation

By extracting monthly electric utility billing information for Drexel we have proposed an
aggregation scenario that isolates the savings from aggregation. The following
information contains the most current demand data for one year of Drexel's metered
building groups as found in Table 5.1. The intent is to illustrate more clearly the
comparison between this and other case study sites. Our procedure was to evaluate the
monthly kW demand peaks for each meter and to compare this value with the actual peak
53
of the aggregated load. The difference is referred to as the kW demand reduction. By
multiplying this value by the incremental cost of demanding one less kW, we determined
the approximate savings for demand reduction only. For this case study site, we used
$25.00/kW. Note that this value may not be the same for all cases, nor for all rate
structures, but it allows the annual savings to be compared for each case study site.
In cases where monthly values were not available, they were simply discounted from the
coincident and summed demand. Also note that the monthly kW demand reduction was
rounded to within 10 kW of the actual difference. Monthly mean values for the peak kW
demand were determined only for those months where meter data was recorded.
The kW demand reduction was found simply as a difference between these two values.
By multiplying this kW difference by the marginal cost of using one less kW, we were
able to isolate the monthly and yearly savings that would be realized if such an
aggregation were to occur.
Savings reflected for the entire year come to $128,500, not including the reduced
commodity charges or the additional buildings on campus that were not considered as
part of proposed aggregation scenario. This sum is approximately 4.8% of the annual
electric bill for the aggregated buildings. Additional savings of $25,000 per year are
realized from the active pre-cooling of the library.
In summary, Drexel University's aggressive building monitoring and control program

allowed aggregators to closely examine the electric usage profile to identify whether or
not it would enhance their portfolio. Because Drexel has a very favorable profile for this
part of the country (i.e. steady electric usage at night and during cooler months), the
university had an advantage over many other customers considered for aggregation. By
negotiating an agreement with the local distribution company and an electric generator,
Drexel secured a lower monthly bill for the entire campus. The resulting savings were
achieved under the same rate structure and with minor adjustments to building operation.
By controlling multiple buildings in aggregate, Drexel was able to secure additional
savings.
54
Individual Monthly Demand Peak (kW) Monthly Annual Demand
Meter # Jul-99 Aug-99 Sep-99 Oct-99 Nov-99 Dec-99 Jan-00 Feb-00 Ma-00 Apr-00 May-00 Jun-00 Mean fkW) Peak (kW)
1 1,986 1,911 1,815 1,772 1,697 1,493 1,425 1,413 1,436 1,415 1,959 1,905 1,685.58 1,986
2 2,083 2,022 1,902 1,836 1,594 1,566 1,598 1,549 1,488 1,288 1,909 1,974 1,734.10 2,083
3 673 670 672 725 700 628 623 644 621 611 710 662 661.62 725
4 522 539 514 516 477 408 403 395 392 391 572 518 470.62 572
5 234 222 219 200 176 399 241.67 399
6 136 138 140 144 145 268 161.83 268
7 341 343 303 350 171 147 344 372 383 166 402 360 306.83 402
8 709 635 517 748 961 1,212 1,695 1,577 •1,220 794 1,035 1,018 1,010.05 1,695
9 220 212 256 386 293 390 464 453 346 285 500 448 354.42 500
S urn Pecks (kW) 6,534 6,332 5,979 6,333 5,893 6,214 6,912 6,762 6,230 5,271 7,754 6,886 6,425 7,754
CdndctentfkW") 6,081 5,951 5,640 6,006 5,360 5,628 6,314 6,155 5,812 5,059 7,323 6,639 5,997 7,323
Difference (kW) 450 380 340 330 530 590 600 610 420 210 430 250 428.3333333 5,568
Percentcge 6,9% 6.0% 5.7% 5.2% 9.0% 9.5% 8.7% 9.0% 6.7% 4.0% 5.5% 3.6%
DemcndSa/inas $11,250 $9,500 $8,500 $8,250 $13,250 $14,750 $15,000 $15,250 $10,500 $5,250 $10,750 $6,250 10708.33333 $139,208
Table 5.4. Drexel University demand and corresponding savings due to theoretical aggregation.
55
6. THOMAS JEFFERSON UNIVERSITY
Since 1995, Thomas Jefferson University has been involved in a load aggregation
agreement. The incentive for Jefferson is therefore shifted from total building
optimization to total site optimization. Instead of trying to peak shave for every building,
the demand peak of the aggregated profile is blunted by shifting the peaks for each of its
components.
6.1 Site Characteristics
The aggregated load for Thomas Jefferson University is representative of 16 buildings

with a combined gross square footage of 3,527,169. The site is located among the high-
rise buildings of Philadelphia between 8th and 12th Streets. Though the campus is
dominated by clinical, research and laboratory spaces, the building inventory also
includes dormitories, offices, classrooms, a parking garage, a library, and even some
retail space. Because of the urban location, residence halls are not one of the primary
building types on campus, and facilities are required to be packed tightly together. Figure
6.1 shows the layout of the campus, as well as the location of individual buildings.
Building numbers are not in sequential order because they are based upon the owner's
established numbering system for simplified communication. For a description of each
building, identify the metered group associated with the campus map and refer to Section
6.3, which contains detailed information about the entire group and the individual
buildings within it.
56
TlmriCB vMesan Uh\esity
GnpiBlVtp
Figure 6.1. Thomas Jefferson University campus map.
57
6.2 Rate Structure, Utility Profile and Metering
The 16 buildings that form the aggregated load for this case study site do not have
individual pulse meters, but are broken up into seven metered groups as shown in Table
6.1. Electric services include seven separate meters that monitor five dual services2 and
two single services. Thomas Jefferson Hospital has four main buildings that utilize three
of the dual-service meters. The remaining power distribution serves the twelve buildings
of the Thomas Jefferson University. The hospital has an overall load factor of 0.85 while
the university experiences a load factor of 0.65. In addition, each building must maintain
a power factor greater than or equal to 0.95 in order to avoid electric utility penalties. In
cases where this was not previously the case, active power factor correction has been put
into effect, both at substations and in the form of capacitors at every motor over lOkW.
Buildings Total Gross Mean Monthly

Meter # Served Square Footage kWPeak W/ft2
1 1-10-15-22-26 1,104,025 3,720.00 3.369
2 25 778,752 5,213.00 6.694
3 2-3-4-5-6-9 984,836 4,640.25 4.712
4 48 287,286 2,573.21 8.957
5 24 244,109 853.02 3.494
6 37 82,972 494.28 5.957
7 49 45,189 268.78 5.948
3,527,169 17,762.54 5.036
Table 6.1. Thomas Jefferson University metering groups.
Because the W/ft2 across campus is somewhat higher than the normalized demand for
Drexel's campus, we would expect that there would be a greater potential for savings
under this plan. However, this is not the case, as explained in Section 6.5. The reason
for this may be due in part to the metering scheme, which combines many buildings
together, but may also be tied to the diverse usage breakdown for each building. If a
laboratory building already has a significant amount of office space within it, the
advantage of combining its profile with, say, another office building is diminished. In the
case of Meter #1, five already diverse buildings make up over 1 million gross square feet,
representing 30% of the total aggregated area. Also, if there is not significant hourly
diversity among these buildings, then the monthly aggregate peak will more likely
approach the sum of the individual monthly peaks, as seen in Figure 6.2.
2
Refer to Section 5.2 for a description of the dual-service metering system.
58
25,000
20,000
15,000
J*
10,000
5,000
• Sum of Peaks Coincident Peak
0 —i J—
Nov-98 Dec-98 Jan-99 Feb-99 Mar-99 Apr-99 May-99 Jun-99 Jul-99 Aug-99 Sep-99 Oct-99
Figure 6.2. Thomas Jefferson University peak demand of aggregated load vs. sum of
individual peaks.
The information collected as part of our interviews with Thomas Jefferson University has
been presented to show how the campus may be evaluated. Table 6.2 indicates the
monthly electric use in kWh. This chart expresses how much power is actually used by
each metered group, and by the aggregated load as a whole. Also, the monthly load
factor is shown for each group and the entire load in Table 6.3. The mean monthly load
factor is calculated for each meter, and then again for the aggregated load, resulting in an
overall value of 0.74. Note that this value includes an entire year of data.
From the chart of monthly load factors in Table 6.3, the load factors of the first four
metered groups are plotted in Figure 6.3 to illustrate the profile of each group. The dip
for metered group #3 during the month of July is attributed to the fact that a large portion
of the square footage was out of service for that month. The monthly load factors for the
last three metered groups and the aggregated load are shown in Figure 6.4. Again, a
significant dip for metered group #7 may be due to the fact that buildings were only
operated for a portion of the days during the month.
Since the signing of their aggregation agreement in 1995, the campus has realized savings
of 10-15% on its annual electric bills. Savings were due in part to aggregation and with
an additional percentage resulting from a negotiated Economic Efficiency Rider (EER)
discount. With additional equipment and personnel, the university has experienced a
creep in the annual electric energy bill of approximately 3%. However, the total electric
energy bill has decreased from almost $9 million in 1992 to the $6.6 million experienced
in 1999. Fifteen-minute demand data are available for each of the electric services, but
this information must be drawn out of the BAS systems or researched according to
printed utility bills. More careful of analysis of the 15-minute data may reveal why
savings are so substantial when the aggregation appears to be only a small portion of the
cause.
59
Nov-98 Dec-98 Jan-99 Feb-99 Mar-99 Apr-99 May-99 Jun-99 Jul-99 Aug-99 Sep-99 Oct-99
Month
Figure 6.3. Thomas Jefferson University monthly load factor for metered groups 1-4.
Figure 6.4. Thomas Jefferson University monthly load factor for metered groups 5-7 and
aggregated load.
60
Individual Monthly Usage (kWh) Monthly
Meter # Nov-98 Dec-98 Jan-99 Feb-99 Mar-99 Apr-99 Mov-99 Jun-99 Jul-99 Aua-99 Sep-99 Oct-99 Mean (kWh)
1 1,797.849 1,869,822 1,672,002 1,671,489 1,871.802 1,838,025 2,216,007 2,285,181 2,695,221 2,336,526 2,532,231 1,803,168 2,049,110.25
2 2,869.425 2.762.127 Z422.866 2,422,761 2,754,459 2,716,884 3.026,823 3,194,157 3,758,901 3,280,002 3,557,145 2,711.727 2,956,439.75
3 2.317.365 2.277,450 2,021,171 2,015,465 2,270,732 2.220,030 2,669,436 2,592,072 1,396,890 3.099,861 2,665,179 2,393.312 2,328,246.75
4 1.179.331 1,221,595 1,059,525 1,052.465 1,215,810 1,182,861 1,400,762 1,257,192 1,309,761 1.654,049 1,690,841 1.238.501 1,288,557.73
5 321.616 306,917 281,107 276.972 296,793 293,712 330,549 398,925 473,984 402.176 412.520 297.087 341,029.92
6 248.378 273,098 241,024 233.224 258,745 230,420 252,971 265,044 315.512 276,621 299,830 238,082 261,079.02
7 148,998 159,072 141.949 145.060 156,036 146,525 136,061 179,904 161,367 185,702 151,162 155.621.34
Sum 8,882,962 8,870,081 7.839,643 7.817,435 8,824,376 8.481.933 10.043.073 10,128,632 10.130,172 11,210,603 11.343,448 8,833,039 9.367.116
Table 6.2. Thomas Jefferson University monthly electric use.
Individual Monthly Load Factor Monthly

Meter # Nov-98 Dec-98 Jan-99 Feb-99 Mar-99 Apr-99 Mav-99 Jun-99 Jul-99 Aua-99 Sep-99 Oct-99 MeanLF
1 0.77 0.75 0.73 0.69 0.74 0.67 0,78 0.78 0.85 0.76 0.80 0.70 0.75
2 0.79 0,74 0.76 0.67 0.79 0.75 0.79 0.81 0.85 0.78 0.83 0.75 0.77
3 0.76 0.73 0.72 0.66 0.79 0.73 0.71 0.72 0.35 0.79 0.67 0.69 0.69
4 0.80 0.72 0.74 0.66 0.77 0.70 0.66 0.57 0.58 0.69 0.77 0.68 0.69
5 0.54 0.53 0.56 0.50 0.60 0.51 0.49 0.55 0.62 0.58 0.55 0.54 0.55
6 0.74 0.77 0.69 0.70 0.74 0.65 0.65 0.75 0.79 0.72 0.77 0.70 0.72
7 0.80 0.84 0.78 0.77 0.87 0,00 0.73 0.72 0.82 0.76 0.85 0.75 0.72
Coincident 0.78 0.73 0.75 0.68 0.78 0.71 0.74 0.75 0.70 0.76 0.76 0.71 0.74
Table 6.3. Thomas Jefferson University monthly load factor
61
6.3 Metered Group Summary
Metered Group #1 - Thomas Jefferson University
Number of
buildings:
footage:
Primary heating: Steam from University loop.
Primary cooling: 2 x 300 ton variable-speed chillers serve
buildings 1, 9, and 15. Building 10 has
a415-ton chiller for animal rooms and 2 x
1000 ton chillers for remainder. Building
22 included on cooling plant of building
48 and building 26 has no cooling.
peak:
Index:
usage:
Comments: Group 4 picks up the cooling portion of
the electric usage for building 22, but
building number 9 should not be included
in this group. Therefore, the indices
should be relatively close for the actual
characteristics of this metered group.

1 Scott Memorial Library 111,970 Office Library
10 Jefferson Alumni Hall 455,711 Laboratory Circulation
15 Orlowitz Residence 206,891 Housing Other
22 Barringer Residence 160,525 Housing Other
26 TJU Parking Garage 168,928 Warehouse Circulation
7.1% Office 78,884

7.1% 1.5%
1.5% Classroom 16,072
16.8%
0.4% Retail 4,737
1.9% Restaurant 21,439
14.2% Warehouse 156,758
18.8% Housing 207,956 14.2%
1.3% Clinical 14,150
9.3% Laboratory 102,748 14.8%
2.1% Library 23,405
2.7% Theater 29,602
0.6% Data Center 6,189
8.4% Mechanical 93,267 18.8%
14.8% Circulation 163,509
16.8% Other 185,309
2.7%2.1% 9.3% 1.3%
1,104,025
62
Number of
buildings:
8
footage: 7"P>P9T/1,T/ / /,-,.-/// giîpppt^p
Primary heating: Steam from University \
loop.
Primary cooling: 2 x 800 ton chillers with
heat recovery systems. 25
900 ton chiller. 220 ton '¥/
<
chiller serves operating
rooms.
p whhtmimi®fa&
peak:
Index:
usage:
Comments:

25 Gibbon Building 778,752 Mechanical Circulation
11.7% Office 90,978 11.7%

1.1% Retail 8,262 16.9% 1.1%
4.6% Warehouse 35,799
18.7% Clinical 145,689
1.1% Laboratory 8,816
22.2% Mechanical 172,856
20.4% Circulation 20.4%
159,157
18.7%
16.9% Other 131,961
778,752
0.3%
22.2%
63
Number of
buildings:
footage:
V\
Primary cooling: 2 x 500 tons serve building 3. 777777777777777//,
500 and 650 ton chillers serve
buildings 4 and 5. Building 2
has 2 x 180 ton chillers and W7/,
building 6 has its own plant
with 2 x 250 ton chillers, but
typically only one of these
two plants serves both
I
buildings.
peak:
Index:
usage:
Comments: Note that building number 9
is served by the cooling plant
for metered group 1.

2 Medical College 191,789 Office Circulation
3 Foerderer Pavilion 233,930 Circulation Mechanical
4 Thompson Building 171,219 Circulation Clinical
5 Main Building 169,807 Mechanical Circulation
6 Curtis Building 136,566 Laboratory Office
9 Martin Building 81,525 Housing Office
17.2% Office 169,029

0.9% Classroom 8,648 13.4% 17.2%
2.9% Housing 28,868
10.6% Clinical 104,737
0.1% Library 933 22.6%
0.9% Theater 8,617
10.6%
13.4% Other 131,699
17.4%
984,836 0.1%
0.6% 0.9%
64
Number of
buildings: p p ^ ^ l
Total square 287,286 /
footage: / /
Primary cooling: 3 x 830 ton chiller plant serves this group plus
48
Annual demand
building 22.
3,353 kW s
/
/
j
s
4
peak: d
Index:
usage:
Comments: Despite building 22's load, the indices for this
group are unusually high. This may simply be
due to the fact that Bluemle is the primary lab
facility on campus.
No. Building . GSF Primary Use Secondary Use

48 Bluemle Life Sciences Building 287,286 Laboratory Circulation
11.6% Office 33,377 11.6%

1.8% Restaurant 5,218 9.7%
1.7% Theater 5,025 20.1 %
9.7% Other 28,094
287,286 30.9%
18.9%
0.7% 1.7%
65
Number of 1
buildings:
Primary cooling: No information available.
peak:
Index:
usage:
Comments:

24 Edison Building 244,109 Office Circulation
26.3% Office 64,143

4.6% Classroom 11,116
1.3% Retail 3,067 26.3%
1.7% Clinical 4,130
0.1% Theatre 126
21.1%
0.2% Data Center 399
17.8% Other 43,352
244,109 15.6% n 1%
U 1%
0.2% "
66
Number of 1
buildings:
Primary cooling: Not available.
peak:
Index:
usage:
Comments:

37 Medical Office Building 82,972 Circulation Clinical
16.2% Office 13,456 9.3%

16.2%
2.6% Retail 2,123
0.4% Restaurant 345
19.8% Clinical 16,448
8.4% Laboratory 6,974 28.4%
28.4% Circulation 23,531 .:-::-:-:-:-:-:-yi9.8%
9.3% Other 7,704
82,972
12.1% „ _ 1 % 8.4%
67
Number of 1
buildings:
Primary heating: Steam from University loop. I 49 l
Primary cooling:
Annual demand
Not available.
294 kW 7
peak:
Index:
usage:
Comments:

49 Clinical Office Building 45,189 Other Circulation
13.9% Office 6,267

13.9%
1.4% Restaurant 652
0.7% Warehouse 325 26.7%
0.5% Laboratory 224
26.7% Other 12,066 20.6%
45,189
23 8%
10.4%
68
Aggregated Group - Thomas Jefferson University
Number of 16
buildings:
footage:
Primary heating: Steam from University
loop.
Primary cooling: Centralized chiller plants
with minimal DX
serving remainder.
peak:
Index:
usage: Thcrro J#fescnUhvesity
Energy Use 31.87 kWh/sq. ft. QjTjrxslVtp
Index:
Comments:
No. Group GSF Primary Use Secondary Use

- 1 1,104,025 Housing Other
- 2 778,752 Mechanical Clinical
- 3 984,836 Circulation Mechanical
- 4 287,286 Laboratory Circulation
- 5 244,109 Office Circulation
- 6 82,972 Circulation Clinical
- 7 45,189 Other Circulation
12.9% Office 456,134

12.9%
1.0% Classroom 35,836 15.3%
0.5% Retail 18,355
6.7% Housing 236,824
8.7% Clinical 305,476
19.5%
0.7% Library 24,338
1.2% Theater 43,515
8.7%
15.3% Other 539,874
15.4% ^ S j g j g l j a s g ^ 8.8%
3,527,169
0.5% 12% 0.7%
69
6.4 Building Operations and Dynamics
Peak-shaving strategies implemented since 1992 include non-essential system shutoff,
pre-cooling and hot/cold deck resets. Additional plant optimization efforts include a new
chilled water plant installed in 1996 and a campus-wide building automation system
(BAS). While these strategies have the potential to offer demand limiting or shifting
capability, it would seem that under its current contract TJU has concentrated on reducing
consumption.
Additionally, since 1999, a local energy service provider has been contracted to provide
consulting for numerous energy saving projects. Each of these initiatives is evaluated
based on long-term savings. For approved projects such as a comprehensive lighting
plan, variable-speed drives for chillers and air-handlers, and heat recovery systems, the
ESP makes the capital expenditure and recovers these costs as a percentage of realized
savings. Under this performance contracting agreement, the university preserves its
positive cash flow while implementing many new cost-saving measures. Other programs
on the horizon include EMS system installations and upgrades, and a campus-wide
advanced metering system.
6.5 Financial Implications of Aggregation
As explained in Section 5.5, the calculation for kW demand reduction was carried out for
the aggregated load of TJU with the results as shown in Table 6.4 below. For April and
June, the coincident kW peak demand values were not available from the initial utility
information, but were supplied by the university Energy Manager to determine the kW
reduction.
For the $6,500,000 electric bill for the fiscal year 1998-1999, our results show that TJU
would only be able to save 1.5% through aggregation. This figure does not account for
savings from reduced consumption, nor does it account for the other buildings on
campus. Also, the value of $25/kW was used to determine demand savings, which may
slightly underestimate the amount for this case. Because the overall electric bill for the
university has decreased significantly since the agreement was established, we had hoped
to discover large kW demand reductions, but the savings from aggregation itself were
significantly lower than expected.
70
Individud Monthly kW Demcnd Peaks Monthly Annud Demand
Meter # Nov-98 Deo98 Jen-99 Feb-99 Mar-99 Apr-99 Moy-99 Jun-99 Jul-99 Aug-99 Sep-99 Oct-99 Mean Peak (kW)
1 3132 3348 3168 3240 3528 3672 3816 4338 4266 4284 4266 3582 3,720.00 4.338
2 4902 5028 4452 4890 4866 4842 5142 5886 5928 5814 5784 5022 5.213.00 5,928
3 4077 4194 3888 4086 3978 4086 5076 5328 5337 5445 5355 4833 4,640.25 5,445
4 1994 2273 1980 2142 2192 2263 2867 3294 3051 3353 2952 2520 2,573.21 3,353
5 793 784 699 744 690 779 912 1070 1020 968 1010 768 853.02 1,070
6 453 478 482 446 484 474 519 528 536 536 523 472 494.28 536
7 251 255 253 254 249 251 269 281 294 293 294 281 268.78 294
Sum Pecks (kW) 15,602 16,360 14,922 15,802 15,986 16,367 18,600 20,724 20,433 20,692 20,184 17,478 17,763 20,724
CdnactentfkW 15,252 16,257 14,510 15,563 15,688 15,992 18,322 20.110 19,557 20,557 19,989 17,339 17,428 20,557
Difference (kW) 350 100 410 240 300 370 280 610 880 130 200 140 334 4,010.00
Percentcge 2.2% 0.6% 2.7% 1.5% 1.9% 2.3% 1.5% 2.9% 4.3% 0.6% 1.0% 0.8%
Demcnd Sa/ings $8,750 $2,500 $10,250 $6,000 $7,500 $9,250 $7,000 $15,250 $22,000 $3,250 $5,000 $3,500 $8,354.17 $100,250
Table 6.4. Thomas Jefferson University demand and corresponding savings due to aggregation.
71
7. AMERICAN HOME PRODUCTS
This case study examines the aggregation of 23 accounts at 11 independent sites in
Pennsylvania and New Jersey. These locations are all connected by American Home
Products of Madison, New Jersey, one of the candidate sites identified. The unique
aspect of this case study is that the sites were grouped together despite the fact that they
are geographically remote and financially autonomous. This is the closest example to
individual buildings of known profiles being aggregated into a common portfolio.
Drawbacks include the fact that because they are multi-building sites, some amount of
diversity is already taken into account by the metering system, just as in the other two
case studies. However, because different owners operate each site, our hope was that
they would shed light on the topic of multi-building control.
7.1 Background
A local consulting company was responsible for managing the aggregation of these sites
and organizing the proposals from competing providers. The consultant sent out a
Request for Proposal and then assessed the respective bids for supplying and transmitting
power. The total load was broken down into 11 accounts at five sites in Pennsylvania
and 12 accounts at six sites in New Jersey, for a total of 23 separate accounts. In the RFP
and for the purposes of our analysis, each account was considered individually, without
any concern for the fact that several accounts may be located at the same site. The reason
for this assumption is that each of the accounts could potentially have a different rate
structure. Multiple rate structures for the same site can be explained by different building
sizes or uses. Because no square footage information was available for the accounts, we
were unable to determine the effective indices for demand and consumption. However,
the consultant for this project was able to identify the primary use type for each account
as indicated in Table 7.1. This table also shows the annual consumption for each site in
kWh.
As explained by the consultant, the purpose of this project was to secure lower monthly
electric bills for each of the sites making up the total load. Combining the 23 accounts
into one RFP and soliciting bids for the entire block drastically reduced the cost of doing
business reduced, for both the bidders and the consultant. In reality, however, the final
product was not, in fact, an aggregation in load, but simply a bundling of information for
easier processing. In effect, each of the sites signed individual contracts securing a rate
that was independent of all the other sites. Furthermore, each site was responsible for
maintaining essentially the same electric usage profile for the duration of the contract.
Penalties were assessed for any significant deviation from the profile submitted with the
RFP, but these penalties were billed to a specific site rather than being distributed among
all the sites. Finally, at the expiration of the contract, each site was responsible for
choosing to continue or terminate the relationship with the contracted provider.
72
No. Location Primary Usage Annual kW Total Annual kWh Hourly Data
Peak Available
1 Pharmaceutical Manuf. 4,226 21,388,626 Y
2 Pharmaceutical Manuf. 2,498 13,330,000 Y
Vaccine Manuf. 6,458 38,908,800
4 Vaccine Manuf. 58 275,520
Vaccine Manuf. 99 287,600
Pennsylvania Sites
5 Research & Development 756 1,824,700

6 Research & Development 2,508 9,989,439 Y
7 Research & Development 578 1,803,900
10 Food Manuf. 6,644 32,704,026 Y
"11 Food Manuf. 714 4,414,808
12 Pharmaceutical Manuf. 376 2,710,700
13~ " Pharmaceutical Manuf. 101 442,560
15 Pharmaceutical Manuf. 1,216 5,719,200
New Jersey Sites
18" Administrative 1,244 7,732,800

iq Research & Development 7,788 40,741,400 Y
20 Research & Development 131 726,720
21 Research & Development 68 247,620
22 Administrative 2,246 10,190,400 Y
23 Vitamin Manuf. 570 1,709,100
51,680 261,244,569
Table 7.1. Electric profile Information for American Home Products' Sites in New
Jersey and Pennsylvania.
7.2 Information Exchange

To ensure that the bidding was competitive, the consultant assembled information for
each of the sites and distributed it in the form of an RFP. This document contained
monthly consumption data for one full year, including peak kW, total kWh, and the
corresponding load factor for each month. For nine of the sites, the actual load profile
data (8760 points) were available in electronic format. Where such hourly information
was not available, the responsibility for estimating usage was placed upon the individual
bidders. In some cases, bidders were unable to provide any price for service, either
because they did not have service in the appropriate location or because they could not
generate a competitive price from the profile information given.
73
7.3 Bid Assessment
Following the RFP and bid submission, the local consultant evaluated the pricing
structures offered by each of the bidders. The first significant hurdle was breaking the
bids into groups based upon duration of the agreement. While the RFP included profile
information for an entire year, the electric-generation-service providers submitted
proposals for six months, nine months, one year, five years and many variations in
between. Some bidders excluded the summer months because of the risk associated with
market volatility during warm weather. Others guaranteed long-term contracts, but with
the provision that pricing be determined by deducting a percentage from the market price.
In addition to these complications, some providers were only willing to service sites in
one of the two states represented by the group.
In all cases, the consultant was comparing the pricing structures to the "cost to compare"
or "shopping credit" for each site, which is the total annual cost for generation and
transmission divided by the annual energy consumption in kWh. The shopping credit
provides a single rate value that combines energy use, demand and other types of electric
charges such as the customer service charge, tariff adjustments or power factor penalties.
This results in two $/kWh figures, one for supply and one for transmission. While many
of the sites had been subject to different rate structures, the discounts offered by the
bundled tariff agreements were all to be calculated in a similar fashion. They were
determined by a fixed price for firm delivery that was multiplied by the monthly usage in
kWh. The local distributor then issued credits for generation and transmission
accordingly, so that each site was still billed according to the same rate structure as
before. Table 7.2 shows the comparison of three example companies.
The figures in Table 7.2 are for illustrative purposes only. For this reason, all pricing
figures have been changed slightly, and suppliers have been rendered anonymous by
referring to them as Company A, B, and C. Nonetheless, we are clearly able to see how
evaluation of the bids can be complicated. First of all, the shopping credit is not fixed for
a particular site. For example, the six-month shopping credit may be for September
through February while the twelve-month price-to-compare includes the high-priced
summer months. In this case we would expect the twelve-month price-to-compare and
the corresponding bid both to be higher. Company C appears to offer more savings for
site 21, but because it is over twelve months, the savings per month are actually lower
than either of the other two bidders. In fact, Company C was not very competitive on any
of the sites, which begs the question, is there too much risk in a twelve-month bid to
ensure a lower pricing structure? From the choices compared in Table 7.2, we can
clearly eliminate Company C.
74
Six Month Proposals Twelve Month Proposals
Shopping Company A Company B Shopping Company C
Site kWh Credit Price | Savings Price | Savings kWh Credit Price | Savings
12 1,395,900 $ 0.0426 $ 0.0422 $ 557 $ 0.0479 $ (7,385) 2,710,700 $ 0.0429 $ 0.0429 $ (117)
13 227,040 $ 0.0475 $ 0.0410 $ 1,474 $ 0.0476 • $ (30) 442,560 $ 0.0480 $ 0.0468 $ 526
~ '" 14 2,650,000 $ 0.0460 $ 0.0419 $ 10,947 $ 0.0475 $ (3,961) 5,501,200 $ 0.0456 $ 0.0464 $ (4,500)
15 2,749,6001 $ 0.0480 $ 0.0433 $ 13,009 $ 0.0476 $ 974 5,719,200 $ 0.0476 $ 0.0478 $ (1,027)
16 1,669,560 $ 0.0490 $ 0.0422 ! $ 11,325 $ 0.0477 $ 2,200 3,601,160 $ 0.0480 S 0.0489 $ (3,160)
17 10,262,190 $ 0.0478 $ 0.0424 , $ 55,504 $ 0 0415 $ 64,497 22,747,390 $ 0.0464 $ 0.0457 $ 14,916
18 2,054,800 $ 0.0526 $ " 0.0427 ' $ 20,251 $ 0.0482 $ 9,135 7,732,800 $ 0.0483 $ 0.0497 . $ (10,551)
19 18,307,200 $ 0.0448 $ 0.0400 $ 88,512 $ 0.0419 $ 52,804 40,741,400 $ 0.0436 $ 0.0430 $ 24,859
20 391,680 $ 0.0486 $ 0.0452 , $ 1,349 $ 0.0436 $ 1,948 726,720 $ 0.0493 $ 0.0474 • $ " 1,361
21 125,760 $ 0.0541 "$ " 0.0450 I $ 1,140 $ 0.0441 S 1,263 247,620 $ 6.0543 $ 0.0487',"$ 1,377
22 4,843,200 $ 0.0428 $ 0.0406'"$ 10,675 "$" 0.0417 " $ 5,431 10,190,400 $ 0.0460 $ 6.0453 S $ ' "7,117
23 761,000 $ 0.0606 $ 0.0430 $ 13,430 $ 0.0482 $ 9,435 1,709,100 $ 0.0582 $ 0T65I2 ."$ (1,775)
Total Savings $ 228,173 $ 136,310 $ 29,027
Savings per Month $ 38,029 $ 22,718 $ 2,419
Table 7.2. Comparison of three bids by electric suppliers.
75
Next, we note that Company B offered more savings than either of the other two bidders
for sites 17, 20 and 21, but if we look at the incremental difference, it is relatively small.
A total of $9,715 could be saved if we elected to have these sites served by Company B.
The question then becomes, what extra costs are associated with doing business with two
electric service providers? Is it not simpler to select Company A and sacrifice the
additional 4% that could be saved?
In the end, American Home Products did just that. One provider was selected for all of
the sites in both New Jersey and Pennsylvania and American Home Products signed
individual contracts for each of the sites for a period of six months. The agreement was
selected because it offered substantial savings, yet was as straightforward as possible.
Most of the risk was held by the supplier, and at the end of the agreement, each site was
free to examine other agreements. The provider had a good record with AHP and was
large enough to convince the consultant that it would be able to meet the needs of the
entire group, even during peak periods. One drawback to the contracts, however, was
that according to many rate structures, if no independent supplier is designated during
summer months, a site is required to purchase power through the local provider at the
market rate. Since the contracts all expired in early spring, the entire process had to
repeated before summer to ensure continued savings.
7.4 Conclusions
Interestingly enough, what both American Home Products and the local consultant
referred to as an "aggregation" was actually just a bundled package of many sites being
evaluated simultaneously. Bidders had less hassle than if they had gone after each site
individually, and the client did not have to incur the expense of the additional meetings
and consultant fees that would be required for evaluation of those individual bids. Also,
because they offered competitive prices on most of the sites, the selected provider won
out even on those sites where the lowest price was available through another company.
The next step would naturally be for an aggressive company to successfully manage these
many sites, thus taking advantage of the additional savings from a true aggregation.
The most valuable aspect of this case study is that it clearly identified the information
that an energy service provider requires in order to offer the most competitive pricing. In
the absence of hourly data, the supplier is assuming more risk, and therefore, the offer
must have either (1) higher base rates or (2) stiffer penalties for deviations from the
submitted annual profile. As market fluctuations diminish and competition becomes
more aggressive, the potential savings to the owner decrease. In this situation, additional
services and the supplier's performance history become increasingly important. Also, as
in the case of American Home Products, the most inexpensive option may not necessarily
be the one chosen for long-term electric service.
76
8. CASE STUDY SUMMARY
Potential
Metering Total GSF of Demand Average Savings from Percent $
Organization Groups Aggregation Reduction W/ft2 Aggregation Savings
Drexel University 9 2,195,364 5,140 kW 3.019 $128,500.00 4.8%
Thomas Jefferson
University 7 3,527,169 4,010 kW 5.036 $100,250.00 1.5%
American Home
Products 23
Our evaluation of these case study sites was intended to contribute information on how
loads for individual buildings have been aggregated in the past. The control and
management of these multi-building groups is the focus of the report, but our research
indicates that for these examples and all others we have found, demand-limiting
strategies are being currently implemented as a result of aggregation agreements.
Deregulation has lowered commodity prices so substantially, at least in Pennsylvania,
that building and site owners can easily save as much as 10% of their previous annual
electric utility bill. This makes demand-limiting projects of a large scale prohibitive
because in many cases, the projects do not meet the financial payback threshold
established by the building owners. Operational changes have also been shunned because
they often sacrifice comfort or profitability.
77
9. OVERALL DESCRIPTION OF TOOLS
9.1 Background
This chapter marks the beginning of the report on the second of two phases of the project.
Phase I involved the identification and documentation of case studies dealing with the
operation of aggregates of buildings. The cases studies revealed that the motivation for
aggregating multiple buildings under the current state of affairs was essentially energy-
cost related (i.e. better rates could be negotiated because of the larger aggregated load).
Very little was being done in terms of load control in either individual buildings or in
aggregated buildings.
The goal of Phase II was to identify tools or capabilities that could assist aggregators to
better operate multiple buildings, with the case studies from Phase I serving as typical
strategies that could be used as initial starting points. Because the case studies did not
reveal any strategies for operation of buildings, the principal investigators undertook to
identify the types of capabilities or tools which aggregators are likely to need in the
future. These tools were largely identified based on numerous discussions with energy
professionals with special expertise and working knowledge in this general area, as well
as close input from the PMSC members, many of whom were knowledgeable in this area.
9.2 Current and Future Scenarios
Kahn (1988), although somewhat dated, provides an excellent overview of how the
electric utility industry moved from a growth-oriented environment into a period of
instability, competition, and re-structuring. Many specific examples are used to first show
how the industry worked when conditions were favorable, and then highlight the
complications that arose when the environment turned hostile. The book focuses on the
broad strategic, planning and organizational issues faced by the utility industry. The book
predates, and therefore does not address, the specific issues associated with utility
deregulation. However, it tackles such issues as marginal and avoided costs, independent
power production, demand side utility programs and comparison to other industries (such
as telecommunications).
A book edited by Thumann (1999) is a compendium of several pertinent articles by

professionals in this field, which provide practical advice to people responsible for
purchasing energy. Lessons learned from deregulation in Scandinavia and England are
presented, as well as an update of retail competition in the US, broken up by individual
states. The role of the procurement professional, how energy legislation is likely to
impact power marketing transactions, current happenings in electric utility deregulation, a
10-steo program to successful utility deregulation for building owners, background on
power pools, as well as experiences with large scale load aggregation and power purchase
78
bidding for state facilities are some of the chapters contributed by experts in these areas.
This book remains relevant despite the very recent changes in the deregulation arena.
The unfolding vista of utility deregulation is leading to five distinct types of players,
replacing the one entity - the regulated electric utility - that existed in a regulated
environment (see Figure 9.1):
1. Power producers, which own the power generators and whose only
responsibility is to generate the needed power and pump it into the grid.
2. Power exchanges, which provide a short-term or spot market for the sale of
electricity by producers and its purchase by utilities or power marketers to
supplement any bilateral, long-term contracts the utilities or marketers may have
with producers.
3. Local distribution companies (investor-owned or municipal), which physically

own the power grid used to transmit and distribute electricity (electric wires,
poles, switching stations, and transformers) and which have responsibility to
maintain this infrastructure and serve customers. Local distribution companies
typically charge a flat rate for transmission and distribution, although congestion
pricing for transmission will be used in the future.
4. Power marketers, which contract to provide power to customers who choose not
to buy it from the local utility.
In a regulated electricity market, the local electric utility owned and maintained its own
generators and provided an integrated service to its customers (scenario 1 in Figure 9.1).
79
Bk-
Local
Power
Producers
Companies Power Marketers
J L
Power
Exchanges
Load Consultants
I
Aggregators
Energy Service Companies
Energy Management
Offering Pro- Load
Software Developers Active Load Consultants
Control Companies Management
i n r
Customers
Figure 9.1. Various scenarios and players in the generation and consumption of
electricity in buildings.
In the early days of deregulation, a new player emerged - the load consultants (scenario
2 in Figure. 9.1). Because customers often lacked the knowledge to negotiate intelligently
with power producers, they relied on services offered by local load consultants. Such
consultants usually catered to larger customers, which either had one building or several
buildings that may or may not have been dispersed geographically. The load consultant
essentially performed three tasks as shown in Figure 9.2:
1. Pre-screen the customers based on certain pre-determined criteria to determine

whether the customer is a potentially "good" client,
2. perform a load analysis based on the customer's historic data, which could
involve monthly billing data or hourly monitored data (if available, and if
warranted), and
3. negotiate a contract with a Load-Serving Entity (LSE) on behalf of the customer
in order to procure the needed power at the most favorable price.
80
Individual
Contract
Customer Load Analysis
Negotiation
Pre-screening
Figure 9.2. Services currently offered by load consultants.
The term "load aggregators" when used in the current atmosphere really means load
consultants. It is clear from the above discussion that currently load consultants act as
customer representatives who represent the interests of the customer(s) and negotiate
with power marketers to get favorable rates due to bulk supply-side purchase. Their
involvement is thus limited to the pre-contract stage, and they have minimal (or no
involvement) once the contract is signed.
As utility deregulation matures, these consultants are likely to become more pro-active
and provide active load management services to customer(s) in order to better manage
and control load. It is best to conceptualize pro-active load management as "load
control in response to variable rate pricing" (Gabel et al. 1998). Although this is a form
of demand side management, it is wise not to use this terminology. About 20 years ago,
demand side management was a hackneyed term, referring to any retrofit or operational
practice that a building owner could perform to save energy in his facility (Lee and
Craven 2001). Usually these operational practices involved at best some sort of passive
load curtailment during the peak hours, by installing thermal storage systems, demand
meters in certain equipment (such as chillers), and EMCS systems for lighting load
management.
The shift towards pro-active load management is being driven by market forces. Electric
utilities have strong incentives to offer real time pricing (RTP) rates to encourage pro-
active demand energy management by customers. Peak electric load reduction decreases
costs for generating, transmitting, and distributing power (Gabel et al. 1998). Further,
RTP shifts some of the risks historically born by electric utilities (who hedge against
them by building in safety margins in the quoted pricing) in delivering the needed power
to the customers. Lower electricity charges when averaged over the year are the
incentive to customers who accept shouldering some of the risks during periods of high-
priced spot-market power.
A major challenge in achieving potential RTP savings is related to building controls. A

large commercial building typically has dozens or hundreds of heating, cooling,
ventilating, lighting, and plug loads whose control set points, start/stop schedules and
other operating parameters need to be properly managed to take full advantage of RTP
rates (Gabel et al. 1998). As a result, energy managers of large buildings are realizing that
proper load control can substantially benefit from an infrastructure of measurement and
monitoring, data processing and analysis, and energy and demand reporting. Such
services have been offered, over the last 15 years or so, by Energy Services Companies
(ESCOs). Further, the importance of this aspect has spawned several Energy
81
Management Software developers who are able to offer such comprehensive services,
either on a local basis or remotely through web-based platforms. In parallel with this
trend, control companies are moving towards networking and interconnectivity of all
HVAC equipment (as well as such other services as fire safety, building access, lighting,
elevator control). This effort is spurred by the ongoing development of such open
protocols as B ACnet and LonTalk. These trends indicate the future emergence of Local
Aggregators who would combine one or all of the above services (scenario 3 in Figure
9.1) and provide pro-active load management in the building or facility. This research is
specifically targeted towards load aggregators who would like to offer services involving
pro-active control of a group of buildings (rather than a single one).
9.3 Objectives and scope
The overall objectives of Phase H of this research are:
(a) to identify situations and conditions under which aggregating individual building
loads is attractive in managing total, multi-building load,
(b) to identify and justify the types of capabilities or tools (including analysis tools as
well as technology tools) that will be useful for local aggregators offering active
load management services as utility deregulation matures, and
(c) to identify operating and control strategies for use in individual buildings which
will reduce energy operating costs at the aggregate level.
The scope of this research is not to develop the tools to the extent that they can be used
by concerned load aggregators. Rather, this research will describe the functionality of
these tools in a conceptual manner and provide relevant review of prior work and what
future developments are needed for each of these tools. Due to economies of scale, the
scope is limited to larger commercial customers and not residential and small commercial
customers.
9.4 Simple Example of Benefits of Multi-Building Aggregation
A simple example with two hypothetical buildings will serve to illustrate the benefits of
aggregating diurnal loads. Two office buildings are on a time-of-use tariff with an
energy charge of $0.06/kWh and a demand charge of $10/kW. One building is
predominately south-facing and the other west-facing, due to their placement in an urban
block. In summer, the peak demands of both buildings are strongly influenced by air-
conditioning loads. The peak electrical load for the south-facing building will occur
earlier in the day than that for the west-facing building. If each building is individually
controlled, the objective would be to minimize the individual peaks. If the buildings are
aggregated, the objective would be to minimize the coincident peak that is bound to be
lower than the sum of the peaks of both buildings. Thus, the benefits due to multi-
building load management are essentially similar to those encountered in the individual
building multi-zone case.
82
A more elaborate example to illustrate these benefits will be provided in Chapter 16.
9.5 Issues Pertinent to Building Owners and Load Aggregators
This section briefly lists some of the pertinent issues and concerns from the perspective
of individual building owners on one hand, and pro-active load aggregators on the other.
9.5.1 The Building owner's perspective
(a) Load aggregation

• How much am I going to save?
• What is my financial benefit in aggregating with several other
customers?
• What are the risks involved?
• How will I be affected if my load shape and magnitude change in
the future?
(b) Building dynamic control

• What is the extra financial benefit?
• What type of load control measures will I have to implement?
• What would they cost to implement/ perform?
• How much prior notice will I be given to initiate control measures
(1 hr... 24hr.)?
• What type of load information will be communicated to me and
how (phone, fax, email)?
• How will I be financially remunerated?
9.5.2 From the perspective of the pro-active load aggregators
(a) Analysis tools

• What is the added expertise that I need to hire in order to offer
such services?
• What is the additional investment I need?
• What are my financial gain and the associate risk?
• How do I initially define and cluster potential customers (building
load, building type)?
• What level of analysis sophistication is needed to define my
portfolio of customers (typical days of each month, or whole
year)?
• How do I identify the optimal initial set of customers?
• How can I add or remove customers from this aggregate?
83
What are the total group savings and how to distribute savings
among individual customers?
What types of data and tools do I need to identify optimal control
measures specific to each customer on an ongoing daily basis?
How can I safeguard myself against risk (due to both load and
electric price signal variability)?
(b) Interaction with individual buildings

• What type of metering infrastructure is in place, and what do I
need to invest additionally?
• What type of costing infrastructure is in place to bill customers?
• How do I implement real-time communication for transmitting
desired control measures and associated financial benefits?
(c) Interaction with power pools

• How do I establish a line of constant communication (regarding
hourly rates, amount of load, time horizon)?
• How will the power pool measure and remunerate me for load
control measures that my clients implement?
• Will they buy excess power from my long-term bilateral contracts
if required?
• Will I be able to sell excess power if required?
• What line of credit do I get?
9.6 Tools Needed by Local Aggregators Offering Pro-Active

Load Management Services
We have identified 11 analysis and technology tools that will be directly useful to a load
aggregator. Some of these tools exist in some form or another, while a few are totally
new). The functionality of these tools is elaborated in Chapters 10 - 15. A brief summary
of the various tools follows Figure 9.3, a schematic diagram of their organization. The
first set of four tools concerns assembling an aggregate of customers. The next set of
four focuses on the details of managing load and the last set of three tools concerns flow
of information and bookkeeping.
84
Tool 2:
TooM: Tool 3:
Assessing Customer
Customer Portfolio
Load Reduction
Pre-screening Optimization
Potential
Tool 4:
Contractual Tool
1
' V i
Tool 7:
Tool 5: Tool 6:
Interaction with
Load Comfort Penalty
Local Utilities/
Forecasting Framing
Power Marketers
i '
Tool 8:
Load
Optimization
''
Tool 11:
T00I9: Tool 10:
Financial
l i n y tx
Remuneration
Custo mers Verific ation
Determination
Figure 9.3. Flowchart of the various tools needed by load aggregators offering pro-
active load management services.
Tool 1: Customer Pre-Screening
The objective of this tool is to be able to pre-screen a customer for inclusion as a viable
candidate for an aggregation portfolio. In practice, customers might need to satisfy a
number of criteria in order to merit further consideration. Criteria include the magnitude
of energy bills, building type, electricity rate structure, electricity load factor, building
operating schedule, metering and communications infrastructure, whether the building is
occupied by the owner or tenants, and building usage (for example, education, health care
or high-technology, which may be criteria for membership in a power-purchase
consortium).
85
Tool 2: Assessing Customer Load Reduction Potential
A pro-active load aggregator cannot afford to spend a lot of time performing building-
specific simulations for a customer who is only at the screening stage. However, it is
important to assess how load can be reduced. Some of this assessment can be binary in
nature: the presence or absence of addressable lighting control circuitry, for example.
Other information can be determined via relatively simple calculations of the impact of
changes in space temperature set points or HVAC set points, for such common systems
as constant and variable air volume. The proposed tool provides estimates of load
reduction via space temperature and HVAC set point control.
Tool 3: Portfolio Optimization
The objectives of this tool are to determine which of the viable customers should be
aggregated to maximize savings and to perform a risk analysis, i.e., analyze the
soundness of the decision in the face of uncertainties and variability in the available data.
Savings can be due to reducing diversified demand by combining loads that peak at
different hours or by active load-control measures. The essence of this tool is a
comprehensive analysis of the benefits of aggregating subsets of a large group of
potential customers and an analysis of the benefits (positive or negative) of adding a
customer to an existing base.
Tool 4: Contractual Tool
This objective of this tool is to allow different types of contractual rates and clauses to be
evaluated between (i) the aggregator and the customers, and (ii) the aggregator and the
power marketers. Ideally, this tool should be sophisticated enough that the aggregator can
offer different prices and terms and conditions to different customers based on the latter's
volume and ability to curtail load.
Tool 5: Load Forecasting
A pro-active load aggregator needs to forecast loads at several levels: end-use loads for
individual customers, whole-building loads for individual customers, and aggregated
loads. Estimating uncertainties in these forecasts is part of the aggregator's risk-
management effort.
Tool 6: Comfort Penalty Framing
Rather than treat thermal comfort as a hard constraint (minimal allowable thermal
discomfort), this tool proposes that thermal comfort be quantified at the zone, building
and site levels, via formation of a weighted sum of indices computed at the lowest level.
Also important is a log of duration of periods of thermal discomfort. Facilities managers
can then decide the extent to which thermal comfort should be traded against operating
costs.
86
Tool 7: Interaction with Power Marketers and Independent System Operators
Tool 7 concerns price forecasts that can be obtained from power pools or from
independent organizations. These forecasts are needed in order to give customers an
incentive to implement load-control measures.
Tool 8: Load Optimization
The objective of this tool is to find the optimal combination of load curtailment measures
to implement that can meet the required load reduction while minimizing the penalty
function (i.e., an estimate of the percentage of building occupants who are dissatisfied
with their thermal environment). This tool should also be able to perform a risk
evaluation, i.e., perform a sensitivity analysis on the optimal solution obtained.
The proposed tool makes use of a search engine and a building-energy simulation
package. In practice, this tool can be enormously complex, because optimization needs
to account for demand charges that are based on peaks during a month or, for ratcheted
rates, during a 12-month sliding window.
Tool 9: Interaction with Customers
Tool 9 concerns communications between an aggregator and its customers, or between a

power marketer and the aggregator's customers who are served by the power marketer.
Tool 10: Monitoring and Verification (M&V)
This is the basis for proper financial remuneration (the next tool). In order to avoid
disputes, it is essential that adequate M&V methods be in place. These require: (i) that
the necessary measurement hardware infrastructure is in place, (ii) that the necessary
measurements are taken at the right frequency and stored properly, and (iii) that accepted
data processing, analysis methods and standards are being followed to "measure" the
demand (and/or energy) reductions and not rely on empirical engineering judgment. The
final objective is to determine (i) which customer has reduced his load, and (ii) by how
much, after being notified.
Tool 11: Financial Remuneration Determination
Just as utility load-control programs of the past have included provision for paying
participating customers, a program to control loads in aggregates of buildings also
requires a method for allocating the benefits of effective load control and penalizing
ineffective action. Especially important is the need to determine a baseline for payment.
87
10. TOOLS TO DEVELOP AN AGGREGATION
PORTFOLIO (TOOLS 1 - 4)
The first set of four tools - customer pre-screening, assessment of load-reduction
potential, portfolio optimization and aggregator-customer contracts - is proposed to aid an
aggregator in assembling a portfolio of customers. To this end, the tools are intended to
minimize calculations that require extensive building-specific information. For example,
the tool to assess load-reduction potential, Tool 2, should rely on guidelines stemming
from studies of generic buildings or first-cut, simplified calculations of a specific
building under consideration.
10.1 Tool 1: Customer Pre-screening

10.1.1 Objective
The objective of this tool is to pre-screen a customer for inclusion as a viable candidate
for an aggregation portfolio. In practice, customers might need to satisfy a number of
criteria in order to merit further consideration, as noted below.
10.1.2 Analytical Criteria
The following list has been assembled based on discussions with load consultants and
aggregators:
• Total annual energy cost - this should exceed a certain value for it to be cost-
effective to the load aggregator to include the customer in an aggregation
portfolio.
• Building type - this factor provides a preliminary indication of the types of loads,
their diversity, and how the building is operated both diurnally and seasonally.
• Building size - size should be in excess of a certain value.
• Building age and recent upgrades - these indicate the state of the mechanical
equipment.
• Current rate structure - this critical factor dictates to a large extent the cost
savings associated with aggregating a given customer.
• Load factor - this is also a critical factor because bilateral or spot power markets
may favor users with flat diurnal profiles. Hence, those buildings which have flat
load factors or loads outside of the peak daytime hours will tend to be given
special consideration.
• Operating schedule - this includes diurnal, weekday/weekend, and seasonal
schedules. Of special relevance is whether the building is partly occupied in
summer when electricity peak rates are high.
88
• Owner occupied or speculative - this aspect is likely to affect load control
strategies that impact visual or thermal comfort.
• Ability to curtail load - this is also a critical factor because buildings which such
ability will be priority customers to load aggregators.
• Membership in a certain class of buildings, such as those operated by a
municipality, public or non-profit facilities for health care or education, or those
owned and operated by high-tech companies.
10.1.3 Other requirements (metering and current rates)
Along with the above analytical criteria, there are other factors that need to be taken into
consideration:
• Physical location of building - there are geographical areas with very high
electricity demand, where the existing electric grid infrastructure is close to or has
reached its maximum capacity. Transmission-system users (those who have
purchased power on the bilateral or spot market) may be assessed a transmission
fee that depends on the degree of transmission congestion.
• Facilities infrastructure - this factor is important because a load aggregator and an
energy manager (and/or the facility technical staff) must communicate frequently.
• Degree of existing building system interconnectivity and automation - this is
important because load aggregation measures must be implemented
automatically.
• Energy charges in excess of a minimum threshold.
• Presence of peak - demand-recording meters, or continuous (i.e., 15-minute
interval) load-recording meters.
• Presence of a data-recording network for centralized access to electricity-usage
data, in either real-time or archival form - this is essential for verifying savings if
load management measures are implemented.
Currently, load aggregators base their pre-screening decisions on experience rather than
any formal analysis procedure.
89
10.2. Tool 2: Assessing Customer Load Reduction Potential
10.2.1 Objective
The objective of this tool is to create a knowledge base containing load reduction
potential (in kW) and the associated penalty or discomfort index in each building under
consideration by an aggregator.
10.2.2 Different Methods of Reducing Loads

There are three types of curtailable loads:
1. On/Off (Lights, elevators and equipment)
This includes equipment that is either on or off (binary type).
a. Optimal Run Time (Zajac 1997)

This scheme involves operates the equipment based on the ideal times for start
and stop that can still satisfy the conditions required during the occupied periods.
This criterion allows the building temperature to float during unoccupied periods,
which reduces energy demand by allowing the equipment to operate at partial
loading.
b. Load Rolling
Also known as load or duty cycling, this scheme reduces energy consumption by
stopping certain electrical loads when they are not required, such as fan operation.
This operates on the principle that the building HVAC system is designed to
operate at a maximum load, but during times of normal loads not all of the
equipment is required.
2. Ramped (Motors and HVAC equipment) - Hackner et al. (1984), Zajac (1997)
This includes equipment for which the electricity peak can be reduced by
changing set point or by using dimming switches for lights. Strategies include
deck-temperature reset, slowing down fans or pumps, reducing ventilation air,
chiller set-point reset and chiller demand limiters. Many of the strategies could
and should be implemented as part of an optimal control scheme that minimizes
HVAC energy consumption (or demand or cost, depending on the objective
function) by adjusting set points.
a. Reduction in supply pressure set point in VAV systems

Pressure can be reduced to the point where VAV terminal boxes in one or more
zones are wide open, thereby saving fan power.
90
b. Slow roll fans or pumps
The speed of fans and pumps equipped with variable-speed drives can be reduced
to a fixed level below what the control loops would have required.
c. Demand limiting
This feature is dependent on the electric company's type of billing. Often, electric
companies charge a demand fee, which is recorded in specified time increments,
in addition to the usage fee. This feature spreads out the demand over the course
of a day in order to keep the demand curve constant, which will result in a lower
demand fee. Although some of the equipment may have to be operated for a
longer period at a later time, limiting electric demand below a pre-set limit to
achieve a flat demand curve can still reduce energy consumption. For example, a
chiller may have a higher efficiency when run under low load conditions at night,
as a function of both fractional load and reduced outdoor temperatures.
d. Reduction in ventilation air

Because energy is needed to condition outdoor air, reducing the amount of
outdoor air reduces energy peaks. This is a short-term measure because it can
compromise comfort. It is best implemented in unoccupied areas.
e. AHU supply-air temperature reset

This feature adjusts the supply air temperature into a space based on the loads that
are present in the building. This temperature is based on the warmest or coolest
space in the system, and the change in the supply air temperature reduces the load
on the coil.
f. Chilled-water temperature reset

This feature adjusts the chilled water temperature to the highest point possible,
while still satisfying all of the spaces. This reduces the load on the chiller and
reduces energy consumption. This feature can also be applied to the heating
water system and boiler.
g. Condenser-water temperature reset

The temperature of the condenser water, which is used to reject heat from the
chiller, can be reduced in order to increase the performance of the chiller.
Although reducing the condenser-water temperature will result in greater energy
expended in the cooling tower fans, this is offset by the energy saved in the
operation of the chiller.
h. Chiller sequencing
Chillers are designed to operate within a certain efficiency range (usually 40-
90%), and this feature takes advantage of these properties. Sequencing causes the
chillers to operate within their designed efficiency range by using more than one
chiller when the operation of only one would cause it to be outside of its
efficiency range.
91
3. Thermal Storage (Dynamic)
This includes using the heat capacity of certain devices to reduce peak loads. A
common method is to use ice storage systems (Kawashima et al., 1996, Henze et
al. 1997). A more recent method is to use the thermal capacity of the building
mass and furnishings (Braun 1990, Andersen and Brandemuehl 1992, Braun and
Chaturvedi 2002). However, these approaches are outside the purview of this
research, and we shall deal exclusively with (1) and (2) above. However,
curtailment of binary loads or reductions in HVAC service reduce either heat
added to a space or the removal of this heat from the space and therefore alter the
normal (no control) flow of heat into or out of building mass. Estimates of load
reduction may therefore need to account for the impact of mass.
10.2.3 Approaches to Determining Load-Reduction Potential
It is desirable, and perhaps necessary, to determine the magnitude of allowable load

reductions, the permissible duration of such load reductions, and how both the magnitude
and the duration are influenced by real or perceived impact on thermal and visual comfort
and occupant performance, and the willingness of the building manager to communicate
to occupants a need to curtail load. Load-reduction potential can be determined by:
1. Measurement
2. Semi-empirical methods
As part of ASHRAE research project 833-RP, Norford et al. (1998) developed a

method for estimating load-reduction potential that relied on a rule base, simple
equations, and interaction with a building operator. For example, lighting
curtailment required a characterization of a space in terms of availability of
daylight, duration of use (short, as for a hotel corridor, or sustained), importance
of use, and availability of dimming controls.
3. Simulation
Such energy-analysis programs as DOE-2 are well suited for relatively precise
estimates of some types of load reduction. However, considerable effort is
required to set up a DOE-2 simulation and to interpret the results. Reductions in
electrical lighting as permitted by daylighting are well within DOE-2's present
capabilities. The impact of reduced fan or pump motor speeds is not. Current
modular simulation programs, including TRNSYS (Klein et al. 1976) and
HVACSIM+ (Clark 1985, Clark and May 1985, Park et al. 1985,1986), make it
possible to simulate HVAC systems at the level of individual components and
their control strategy. EnergyPlus (EnergyPlus 2002), the replacement for DOE-2
(DOE2 2002) and BLAST (BLAST 2002), includes this modeling flexibility.
92
Other software generates commercial and residential whole-building and end-use
load shapes for electric and gas customers throughout the U.S., based on historic
and extensive data bases. Examples of such software are an older program,
HELM (EPRI1985), and a new program called PowerShape Version 2.0 (EPRI
2000). These programs are ideal for identifying customers who are profitable to
be aggregated, as well as those who are not. How changes in equipment, operation
and weather affect these load shapes can also be evaluated.
10.2.4. Examples of Estimates of Load Reduction Potential
It is possible to estimate load reduction potential on a building-specific basis with little

information, via the use of steady-state equations for mass and energy balances. These
equations have been implemented in a spreadsheet program, which facilitates changes in
parameters. The steady-state approach includes no thermal dynamics, by definition.
There is no time history of increases in zone temperature as supply-air temperature
increases, for example. In this case, it is assumed that the increase in supply-air
temperature would be in effect long enough for the zone temperature to reach
equilibrium.
It could be argued that load-reduction potential would be better estimated with a dynamic
simulation, such as EnergyPlus. Such a simulation would make it possible to track
changes in zone temperature and how these changes affect load reductions. Demand
reductions could then be averaged over a period of several hours, rather than being
calculated only in steady state. However, the steady-state approach will be shown to give
reasonable results, with the minimal effort that is appropriate for tools to be used before
customers are contractually committed to an aggregator.
10.2.4.1. Increase in zone-temperature set point
Consider the benefit of raising zone temperatures, for both constant-volume and VAV
systems. Implementation of such a measure requires digital thermostats that can be reset
centrally, via the Building Energy Management System. Mass and energy balance
equations can be used to estimate the savings in steady state.
/-\
UAfacade (T
\~ oa
—T )
~ zone t
x . /-\ 1A 1
facade \ oa zone I
zone
where
UAfacade = conductance of fa$ade, W/m 2

Toa = outside-air temperature, °C
Tzone = zone temperature, °C
Azone = area of zone, m 2
Qim = heat gain per unit floor area from occupants, lights, equipment and solar
radiation, W/m2
93
v = T'zom 102
where
Vzone = volumetric flow per unit floor area, m 3 /s m 2

Qzone - total zone heat gain per unit floor area, W / m 2
p = density of air, kg/m
Cp = heat capacity of air, J/kg K
Tsa = supply-air temperature, °C
Qcoil = (Vzone ' V

oa X 1
^ ~TJ)+ V
oa iToa ~ Tsa ) 10.3
where
Qcoii - sensible coil load per unit zone floor area, W / m 2

voa - outdoor airflow per unit zone area, m3//s m2
Chiller power can be modeled with the method developed by Gordon and N g (2000),
who developed the following relationship for chiller coefficient of performance (COP):
(Tcdi-Tcho) _ (l/COP + l)Qch

+1 —I —«j '"2 r«3 IU.*r
COP *•cdi Qch * cdiU-ch * cdi
This can be rewritten as:
Qc ATc , . - r , ) + a.T J +aAT,.-T. ) + a„Q2u

p= " "i cho 1 cdi cho 2 cdi cho 3 en iQ r
cho 3 eh
where
P = electric power k W
QCh = chiller evaporator thermal load (kW)
Tcdi = condenser water inlet temperature (K)
Tcho = evaporator water outlet temperature (K)
Data from a representative 434 Ton centrifugal chiller were used to plot chiller power as
a function of load, evaporator outlet temperature and condenser inlet temperature, as
shown in Figure 10.1.
94
Figure 10.1. Power of a centrifugal chiller as a function of load, evaporator outlet
temperature and condenser inlet temperature.
Equations 10.1-10.3 were used to estimate the reduction in fan power and (constant
COP) chiller power due to an increase in zone temperature. Estimates are building
specific but can be made with minimal information, as shown in Table 10.1. Savings are
shown in Figure 10.2 for a VAV system, in Figure 10.3 for a CAV system, and in Figure
10.4, comparing the two systems. The data are for a specific internal load, 30 W/m2.
Figures 10.5-10.10 provide two more triplets of plots, for 40 and 50 W/m2 internal load.
95
glass U 2.5 W/m2
wallU 0.5 W/m2
WWR 0.5
Awall 30 m2
Afloor 100 m2
rho 1.2 kg/m3
Cp 1000 J/kgK
Qint 30 W/m2
Tzone 22 oC
Tout 30 oC
Tsa 13 oC
max fan flow 5 L7sm2
max fan power 10 W/m2
fract pressure
static 0.4
chiller COP 3
OA flow/m2 1 L7sm2
Table 10.1. Inputs for calculation of energy savings due to an increase in zone
temperature.
96
-fan savings
-cooling savings
Increase in zone temperature, oC
Figure 10.2. Demand reduction due to an increase in zone temperature for a VAV
system, with internal loads of 30 W/m .
I
,2 4
-cooling savings
»- 2
IB
E
a>
Q 0+
Figure 10.3. Demand reduction due to an increase in zone temperature for a CAV
system, with internal loads of 30 W/m2.
Figure 10.4. Comparison of demand reductions for VAV and CAV systems due to an
increase in zone temperature, with internal loads of 30 W/m .
97
-fan savings
-cooling savings
-cooling savings
98
2 6 -,
& 5-
J t ^ * ^
S 4 4
^ * ^ ^ m r - ^ * ^ —•—fan savings
| 3
^ ^ * - —•— cooling savings
2 2-
11 / * ^ ~
v Un J / ^
Q *
(3 1 2 3 '%
CM
I
,2 4
-cooling savings
c
IS
E
0+
1
99
For the VAV system, savings in cooling power and fan power are of comparable
magnitude over a range of internal loads. Cooling savings are greater than fan savings
for the smallest internal loads, when less airflow is needed and fan power is lower. For
the highest internal load, savings in fan power exceed those for cooling. Demand
reduction associated with cooling increases with internal load for a given increase in zone
temperature, because fan power increases and cooling is required to remove fan heat from
the air stream. In contrast, savings for the CAV system are confined to cooling and, as
expected, are independent of internal load.
The demand reductions for an internal load of 40 W/m are similar to those reported by
Haves (2002) in a study performed at Lawrence Berkeley National Laboratory (LBNL).
Haves used the DOE-2 building-energy simulation program to study prototypical
buildings in five California climates; his estimate of savings due to an increase in zone
temperature was independent of climate. His demand reductions were averaged over the
four-hour period from 2 p.m. to 6 p.m.
The key result here, shown clearly in Figures 10.4, 10.7 and 10.10, is that savings from
an increase in zone temperature will be substantially larger for a VAV system, where fan
power will vary with flow, than for a constant-volume system. To evaluate potential
clients, an aggregator should first determine whether the zone set points can be centrally
programmed and then look for a VAV rather than constant-volume system.
10.2.4.2. Reduction in lighting load
The same equations can be used to evaluate a reduction in lighting load and associated
reductions in fan power (for a VAV system) and chiller power. Lighting reductions
under real-time pricing have been reported by Gabel et al. (1998). These reductions
were made in specific lighting circuits in a hotel. For office applications, lighting
reductions are difficult to achieve because lighting panels are dispersed and control costs
are currently high. As the cost of addressable electronic ballasts drops, it will become
possible to communicate with lights via smart lighting panels that can act as a gateway
and can in turn communicate with a BEMS via a BACnet or other interface.
Results are shown in Figures 10.11-10.13, for a reduction in total internal gain from 50 to
40 W/m2. These results apply only to fans and cooling power and are in addition to the
demand reduction associated directly with lights. For example, Figure 10.11 shows that a
reduction in lighting power of 5 W/m2 will lead to additional savings of about 2 W/m2 in
both fan power and cooling power.
Demand reduction for the VAV system is moderately smaller if the reduction in internal
gains spans 40 to 30 W/m2, because fan power is sensitive to the magnitude of internal
loads. Savings for the CAV system are limited to cooling and are independent of the
internal-load reference point (50 or 40 W/m2). The total savings for a VAV system are
more than twice that of a CAV system, making the former more attractive to a load
aggregator if lights can be controlled.
100
w
5
3
-•—fan savings
fa -m— cooling savings
1 ~ ^ ~
I o l-^T
0 2 4 6 8 10
Decrease in lighting load, W/m2
Figure 10.11. Demand reduction in fans and air conditioning in a VAV system due to a
decrease in lighting load.
-cooling savings
Decrease in lighting load, W/m2
Figure 10.12. Demand reduction in fans and air conditioning in a CAVsystem due to a
Figure 10.13. Comparison of demand reductions in VAV and CAV systems due to a
101
10.2.4.3. Increase in supply-air temperature
The steady-state mass and energy balance equations can identify changes in fan and
chiller power due to a change in supply-air temperature in a VAV system and cold-deck
temperature in a CAV system. Results vary with internal load and are shown in three
groups of five figures. Figures 10.14-10.18 are based on internal loads of 30 W/m2,
Figures 10.19-10.23 are based on 40 W/m2, and Figures 10.24-10.28 are based on 50
W/m2.
Increasing the supply-air temperature in a VAV system is problematic. As the supply-air

temperature rises, airflow increases to remove a given amount of heat from the building.
Fan power increases and cooling power rises as well, due to the increase in heat
generated by turbulent dissipation in the air stream and by the fan motor. The simulation
does not account for possible reductions in chiller power when the cooling load is
constant, due to increasing the chilled-water temperature as the supply-air temperature
rises. When the VAV terminal boxes are wide open, airflow no longer increases, space
temperatures rise, and cooling power drops.
For the lowest internal gains, the supply-air temperature in the VAV system must
increase substantially, from 13 to 16.5 °C, before the airflow reaches a maximum value.
As a result, fan power rises sharply. Savings due to reduced cooling load are much
smaller than the increase in fan power and the strategy is unsuccessful. Figures 10.13
and 10.14 show the changes in fan and chiller power and the increase in zone
temperature.
Increasing the cold-deck temperature in a CAV temperature leads to no fan-power

penalty. As the cold-deck temperature rises, more of the airflow goes through the cold
deck. When all air is flowing through the cold deck and is not adequate to meet the
cooling load, the zone temperature rises and cooling power drops, as shown in Figures
10.16 and 10.17. The comparison between the VAV and CAV systems, Figure 10.18,
clearly favors the CAV system.
As the internal load increases, demand reduction is less compromised by an increase in

fan power. Airflow and fan power are closer to their maximum levels before the supply-
air temperature is increased. This effect can be seen for an internal load of 40 W/m2 and
is even clearer for 50 W/m2, for which Figure 10.28 shows that the VAV and CAV
systems produce nearly identical reductions in demand.
The lesson here for a load aggregator is to steer clear of changes in supply-air
temperature in a VAV system unless they are implemented on days when building load is
near its maximum.
102
* # ^
9 10
-fan savings
-cooling savings
Increase in supply-air temperature,

oC
Figure 10.14. Demand reduction in fans and air conditioning in a VAV system due to an
increase in supply-air temperature, with internal loads of 30 W/m .
2 8
o -
o
^
2 26-
«3-»
n
| 24-
^y —•—Tzone
£
CD ^ ^
« 22,
c
o
N 20 -
() 2 4 6 8 10
Increase in supply-air temperature, oC
Figure 10.15. Increase in zone temperature due to an increase in supply-air temperature

for a VAV system, with internal loads of 30 W/m .
103
CM T -,
J
E
c 2 y
.0
u
% 1 ^S^_ —•—cooling savings
i o, y
to
E <) 1 2 3 4 5 6 7 8 9 1 0
a> ,
• "'
Increase in cold-deck temperature,
oC
Figure 10.16. Demand reduction in air conditioning in a CAV system due to an increase
in cold-deck temperature, with internal loads of 30 W/m2.
Figure 10.17. Increase in zone temperature due to an increase in cold-deck temperature

for a CAV system, with internal loads of 30 W/m .
Figure 10.18. Comparison of demand reductions in a VAV and a CAV system due to an
increase in supply-air or cold-deck temperature, with internal loads of 30 W/m .
104
-•—fan savings
-9-^ -•—cooling savings

oC
increase in supply-air temperature, with internal loads of 40 W/m2.
30
o
o
28
«S 26
8. -•— Tzone
E 24
o 22 4 • •
o
N 20
0 2 4 6 8 10

for a VAV system, with internal loads of 40 W/m2.
105
CM
E
£
jo
O
3
•o
-cooling savings
•D
C
(0
E
a>
a
oC
Figure 10.21. Demand reduction in air conditioning in a CAVsystem due to an increase

in cold-deck temperature, with internal loads of 40 W/m .
30
oo 28
£
_3 26
"5 -«—Tzone
<B
Q . 24
E
c 22 • • • • • -
o
N
20
0 2 4 6 8 10
Increase in cold-deck temperature, oC

Figure 10.23. Comparison of demand reductions in a VAVand a CAV system due to an

106
CM R -,
-
5 4-
§ 3- mJtM^W^
—•—fan savings
1 *- „***
J * * * * ^
—a—cooling savings
1'
«5 0 1 1***
) 1 2 3 4 5 6 7 8 9 10
S
Q -1' J
oC
increase in supply-air temperature, with internal loads of 50 W/m .
-*—Tzone
0 2 4 6 8 10

for a VAV system, with internal loads of 50 W/m .
107
1 b^
5 4H
1 3,
O 2
3a> - —•—cooling savings
1
£ -
c
<S 0 <
Q -1 J ) 1 2 3 4 5 6 7 R 9 10

oC
Figure 10.26. Demand reduction in air conditioning in a CAV system due to an increase
in cold-deck temperature, with internal loads of 50 W/m .

Figure 10.28. Comparison of demand reductions in a VAV and a CAV system due to an
108
10.3 Tool 3: Portfolio Optimization
10.3.1 Objective
The objective of this tool is to determine which of the viable customers should be
aggregated so as to maximize savings and to perform a risk analysis, i.e., analyze the
soundness of the decision in the face of uncertainties and variability in the available data.
10.3.2 Mathematical Basis
Basically, this tool must analyze the savings by performing all the permutations available.
From probability theory:
n\
Combination of n things taken p at a time = C(n, p) 10.4
p\(n-p)\
However, p can take values from 0 (no customer is suitable for aggregation) to n (all
customers are to be aggregated). Hence:
n
Total number of possible combinations of n items = m = ^ C(n, p) 10.5
p=0
How the value of m increases with n is shown in Table 10.2, and an exponential behavior
is noted. We have identified the following regression model (R-square = 0.999) in order
to determine m more easily:
m = 10A(0.301.n) 10.6
It is obvious that the sum of the possible combinations for even moderate values of n is
very large. For n = 10, the sum m =1,024! Clearly, the pre-screening tool has to be fine
enough to reduce the candidate set to a manageable number. From a practical viewpoint,
the pro-active load aggregator may likely to limit his aggregation pool to, say, 5-10
customers, particularly during the preliminary years.
109
n Sum=m log(m)
0 1 0.000
2 4 0.602
4 16 1.204
6 64 1.806
8 256 2.408
10 1,024 3.010
12 4,096 3.612
14 16,384 4.214
16 65,536 4.816
18 262,144 5.419
20 1,048,576 6.021
25 33,554,432 7.526
30 1,073,741,824 9.031
40 1,099,511,627,776 12.041
50 1,125,899,906,842,620 15.051
Table 10.2 Variation of the sum m of possible combinations with the number ofn items
(following Eq.10.5)
10.3.3 Data Requirements
The most complete load data would be in the form of 8760 hourly demand values (or
even values at 15 or 30 minute intervals). Such data are available from the electricity
provider for larger customers. Alternatively, utility bills are used in smaller buildings.
One also needs to know the electric rate structure or price signal information before and
after aggregation.
10.3.4 Risk Analysis

An aggregator faces uncertainty about the future loads of customers selected for inclusion
in a portfolio. These loads can increase or contract due to changes in business cycles and,
for buildings more subject to facade rather than internal loads, changes in weather.
Energy prices, rate structures, and access to power-exchange market rates can also
change. Sensitivity analysis can be used to assess risk, i.e. evaluate how and the extent to
which the decisions made are affected by uncertainty or variability in the data provided.
Such issues are well known and are described in several textbooks, for example Evans
and Olson (2000), and software tools are available for this purpose (Decision
Engineering, 1999).
10.3.5 Illustrative Example
An example using fictitious building loads and price data will illustrate the capabilities of
the portfolio tool. We shall select four buildings: a school (S), an Office (O), a retail
establishment (R) and a grocery store (G). We have used ELCAP data (Huang and
110
Franconi 1999) to generate typical diurnal load profiles for the four different types of
buildings. These are shown in Figure 10.28. The diurnal electricity rates before
aggregation and the aggregated price signal are shown in Figure 10.29.
Typical Profiles for Four Different Bldg Types
Figure 10.28. Diurnal load profiles for the four different building types used in the
illustrative example.
Ill
Price Signals
S
* 500.00
Figure 10.29. Price signals for the four buildings assumed for the illustrative example.
There are 15 combinations possible with n = 4. (Note that Table 10.2 identifies 16, but
one combination is that of no customers.) The fractional savings, i.e., savings divided by
the pre-aggregation cost for each of the 15 combinations are shown in Figure 10.30,
sorted in descending order. Hence aggregating the School (S) by itself is most
advantageous followed by School-Office (OS) combination, and so on. Figure 10.31
depicts the same combinations but by the dollar savings. Here, we see that S saves about
$75,000 (which is 33% of the price paid by the school). The combination OS saves
slightly over 31%, the dollar savings is slightly over $150,000. If the aggregator wishes
to achieve maximum total savings, he would consider the combination ORGS. However,
the difference between ORGS and ORS is not very important, and rather than take the
risk of aggregating four buildings, the aggregator may wish to aggregate only three
buildings (the set OGS) in order to reduce risk. Performing such absolute evaluations
and assessing corresponding risk levels are the capabilities to be provided by Tool 3.
112
Illustrative Example of Load Aggregation
(4 Bldg Types)
I I n —
•
CO 2 0 . 0 %
-H m
i
i.-
O ORS RS OGS GS ORGS OG OR RGS ORG G RG R

Different Combination*
Figure 10.30. Fractional dollar savings under different combinations of aggregating the
four buildings used in the illustrative example.
Illustrative Example of Load Aggregation

(4 Bldg Types)
$150,000 -|
ORS BS OQS OS ORGS OG OR RGS ORG Q RG

Difforant Combinations
Figure 10.31. Dollar savings under different combinations of aggregating the four
buildings used in the illustrative example.
113
10.4 Tool 4: Contracts between Aggregators, Customers, and
Power Marketers
Load aggregators and their customers require contracts to spell out the nature of the
service and associated charges. Today, as noted in Chapter 9, aggregators are load
consultants or deal makers. They negotiate a contract between their customers and power
marketers that obligates the marketer to provide power but specifies no load-control
activity. Contracts appropriate for aggregators in this role have been developed and this
subject does not require research on the part of ASHRAE or other organizations.
However, contracts that spell out payments for load control and means of defining and
measuring control are not in widespread usage. This section of Chapter 10 briefly
reviews the nature of contracts in use by load aggregators and defines a need for load-
control contracts.
10.4.1 Distinguishing aggregators and load-serving entities
A number of businesses serve groups of energy-consuming customers:
1. Independent marketers;
2. Investor-owned utilities;
3. Municipal utilities;
4. Power marketers associated with investor-owned utilities.
Additionally, large customers (industrial sites, university campuses or hospitals) may

effectively act on their own behalf in seeking power deals from service providers.
It is useful to distinguish aggregators from load-serving entities (LSEs), which include

power marketers and, in some cases, local utilities. The current role of the former is to
put together a deal for its clients and do nothing during the period of the contract. The
aggregator goes out to bid on behalf of its clients, negotiates the best deal, and sets up
contracts between its clients and the power marketer chosen from the bid process. The
power marketer is required to provide power per contract terms and does so on the basis
of bilateral contracts with generators and, as needed, purchases on the short-term power
from markets overseen by Independent System Operators. The tools proposed in this
report are intended to broaden, even re-define, the role of an aggregator.
Independent power marketers currently act purely in the deal-making role defined above.
Because they have an obligation to supply power to customers, they are one type of LSE.
Investor-owned utilities (IOUs) maintain the grid (and in this role can be called Local
Distribution Companies or LDCs) and provide power to those customers who do not
contract with a power marketer. Therefore they also have an obligation to provide power
and are another type of LSE. To the extent that IOUs have sold their generating plants
and must contract for power in long-term or short-term markets, they in effect are seeking
a good deal. The incentive for getting a favorable price for power may rest more with the
114
IOU than its customers, given that the rates the utility charges are regulated and may not
cover purchase costs.
Would an IOU be interested in the tools proposed in this report? The load-cooperative
programs that some IOUs ran in the past, to aggregate small groups of customers capable
of collective action to shed load, would benefit from better estimates of savings potential,
better load forecasts, and better active control measures. Customers could use the tools
directly if given sufficient incentive, in the form of real-time or interruptible rates.
Municipal utilities may be aggregators, narrowly defined, or LSEs. For example, one
California municipal utility maintains some of its own generation and buys power on the
long-term and short-term markets, and is therefore an LSE. Others may maintain the
local grid but negotiate for power on behalf of its customers, in the role of an aggregator.
10.4.2 Independent marketer
One independent power aggregator is the Massachusetts Health and Education Facilities
Administration (HEFA), a public energy-buying consortium in Massachusetts that also
offers energy management services to its members (www.mhefa.org,
www.poweroptions.org, from which web site the following information was obtained).
HEFA's PowerOptions service purchases gas and electricity for its 400 members, which
join voluntarily and pay an annual membership fee based on electricity usage. Members
are given opportunity to review and sign gas and electricity contracts that reflect the best
purchase contracts HEFA can negotiate. Contracts offered by suppliers are evaluated on
the basis of price, reliability, and financial backing.
PowerOptions initially served hospitals, human service organizations, colleges and

universities, and cultural institutions. It was expanded by legislation to offer its services
to any public or non-profit organization in Massachusetts. PowerOptions members
purchase more than $250 million of electricity annually, representing 300 MW of
capacity.
Contracts concern only electricity supply. The local distribution company (LDC) retains
responsibility for transmission and distribution and for metering services. The LDC
issues electricity bills, which include charges for supply, transmission, distribution, and
recovery of assets stranded during the deregulation process.
Rates for transmission, distribution and stranded assets are regulated by the state. For
supply, LDC customers may be on a standard offer or a default offer, or may make their
own deal. The standard offer is a state-regulated offer open to customers who had
accounts as of March 1,1998. Annual prices were established for each LDC. The
standard offer provided relatively cheap electricity at first, but rates have and will
continue to increase substantially over the seven-year life of the offer. The intent was to
give customers an immediate benefit from the deregulation process but later encourage
them to seek a competitive supplier in the retail marketplace. Default service is used by
customers with new accounts (relative to March 1,1998), those who entered the retail
115
marketplace, and subsequently returned to their LDC for supply, and those not eligible
for the standard offer. LDCs bid default service on six-month intervals. The retail
marketplace is expected to offer cheaper service for most customers.
HEFA, through a wholly-owned, non-profit subsidiary, negotiates master agreements

with suppliers to provide electricity under agreed-upon prices and contract terms to
PowerOptions members. HEFA also pre-negotiates a bilateral agreement between the
supplier and the PowerOptions member. This arrangement is desirable if an aggregator
believes that energy prices will rise over time and that a long-term contract will be
cheaper than buying power on the spot market.
The first agreement with an electricity supplier was signed in 1998 and gave members a
choice of three- or five-year contracts with the supplier. A second agreement with a
different supplier was recently signed and is intended for PowerOptions members on
default service. Savings vary across PowerOptions members. The second agreement
offers fixed pricing for smaller accounts, where there is no interval metering. Pricing is
periodically refreshed to reflect changes in the market, but are claimed to be substantially
lower than default prices offered by LDCs. Prices for larger accounts, with interval
meters, are based on market rates and usage patterns, and are currently lower than default
and standard offers. Billing is via a single statement from the LDC or, if requested by the
customer, separate bills from the LDC and the supplier. Contracts under the first
agreement have offered savings in electricity supply (distinguished from transmission and
distribution) of 5-8% relative to the standard offer.
In summary, this form of contractual arrangement makes use of the aggregator as a

negotiator. The aggregator negotiates agreements with suppliers and provides a
contractual template for the use of customers and suppliers.
10.4.2 Standard and default offers from investor-owned utilities
One investor-owned utility (IOU) serving Massachusetts provides location-specific

standard-offer and default-offer electricity (www.nstar.com). The standard offer,
available until 2005, includes location-specific rates for electricity. For example, the rate
for Boston is $0.0495/kWh as of April 1, 2002.
Default service is available on fixed (over two six-month intervals) and monthly variable
rates, as shown in Tables 10.3 and 10.4. Prices for the fixed-rate option are based on the
IOU's contract with the default-service provider. Prices for the variable rate are based on
the actual costs of purchased electricity.
116
Pricing Pricing
Customer Group (1/1/02 to 6/30/02) (7/1/02 to 12/31/02)
Residential Customers 6.393 5.638
Small Commercial/Industrial 6.574 5.671
i
Customers and Lighting
Large Commercial/Industrial 6.549 5.922
Customers
Table 10.3. Fixed-rate prices for default service offered by an Investor-Owned Utility.
Small Large |
Residential Commercial/Industrial Commercial/Industrial 1
Month (2002) Service Rate Customers and Lighting Customers j
January 6.823 7.044 7.346 |
February 6.794 7.105 7.166
March | 6.085 6.311 5.993 _ J
April j 6.034 6.172 5.939
[May I 6.099 6.220 6.023 |
June 6.464 6.536 j 6.715
July | 5.638 j 5.671 7.127 1
August I 5.638 5.671 7.240
September 5.638 j 5.671 5.304 |
October 5.638 | 5.671 4.686 j
November | 5.638 5.671 5.019
December | 5.638 5.671 5.877 j
Table 10.4. Variable prices for default service offered by an Investor-Owned Utility.
10.4.3 Municipal utility
Municipal utilities offer contracts to customers in the form of standard rates that are of
the same form as those offered by IOUs. Rates may include fixed monthly charges,
energy charges that vary seasonally and with consumption, and demand charges (see, for
example, www.smud.org).
Municipal utilities may act as aggregators or load-serving entities. They rely on self-
generated and purchased power. One California municipal utility has served peak
demands as high as 2,700 MW. It has 1,200 MW of its own generation, 57% from
117
hydroelectric sources, 42% form thermal plants (predominantly cogeneration) and 1%
from photovoltaic and wind supplies. As of May 2001, this municipal utility had long-
term contracts in place for nearly 1,200 MW of purchased power, of which the largest
single source was 460 MW from the Western Area Power Administration. Throughout
the year, the utility buys and sells short-term power to meet load and reduce costs.
10.4.4 lOU-affmated marketer
These power marketers function in a way similar to independent marketers, in providing

power to retail customers in regions outside the IOU's service territory.
10.4.5 Contract enhancements to facilitate load control
The bilateral contracts currently in place have no provision for load control. Load control
could be implemented if either the aggregator or its customers, or both, have a contractual
incentive to do so.
At the level of a single building, incentive is most appropriately given to the customer.
As will be explained in Chapter 13, power marketers use power exchanges to either make
up for a short-fall in committed generation during times of high load or to sell excess
power. The power marketer has an incentive to avoid buying high-priced power on the
spot market and an equal incentive to make excess power available. That incentive can
be shared with customers, in an arrangement that would require that the marketer
communicate to the aggregator's customers a forecast of when the marketer anticipates
being short of power. The customers can take that information and a price forecast and
then act appropriately. Customers can be rewarded financially, based on load shift or
curtailment relative to a baseline. The reward can be positive - payment for action - or
negative - a penalty for failing to take contracted action.
The situation is more complicated at the level of aggregates of buildings. Ideally,

aggregated buildings would shed load in a coordinated manner that would minimize the
power marketer's cost of providing service. The marketer would then share the cost
savings. For example, if an aggregate of buildings could, through coordinated action,
achieve a perfectly flat load profile, the power marketer could contract for relatively
inexpensive power from a baseload power plant. If a single member of the aggregate
does not shed or shift load as required during high-load periods, the marketer will again
need to go to the spot market. In this case, it will be necessary to identify which
customer failed to perform and contractually necessary to determine how to penalize that
customer and, conversely, reward all customers for their measured contribution to overall
load shedding. Tool 4 should therefore be a means of evaluating contract incentives for
load control and a set of possible contracts.
118
11. LOAD FORECASTING (TOOL 5)
The objectives of the load-forecasting tool, Tool 5, are three-fold:
1. To forecast for the next 12-24 hours the aggregated load of an individual building;
2. To simultaneously forecast the disaggregated controllable loads; and
3. To estimate the impact of measures that have the potential to reduce the
controllable loads. Load-control measures were described in Section 10.2.2.
This would provide the aggregator the capability of determining whether the aggregated
load is likely to exceed the contractual demand value for the total aggregate, and if so by
how much. How much load to shift or trim and for how many hours during the day could
then be determined.
This chapter reviews work in the literature related to the forecasting of building loads and
concludes that this work is sufficiently broad and thorough that substantial new work is
not needed. What is needed, however, is to evaluate the different levels of approaches
and their associated modeling complexities under different practical circumstances, and
to define guidelines as to their relevance and applicability. The chapter starts with a
review of load forecasts for electrical system loads, which are at a scale much larger than
an individual building but have been predicted with tools appropriate for use in single
buildings. The chapter then continues with work developed explicitly for buildings.
Finally, a simple illustrative example is given.
11.1 Background and Definitions
The term "load forecasting" means different things to different users. There are
essentially three different interpretations to this term, and accordingly three bodies of
knowledge have evolved. These are:
1. Load forecasting using load research data
Electrical utilities have spent considerable time, effort and expense on this activity
from the last three decades under the regulated atmosphere (more exactly after the
1973 embargo). Prior to this time, most Public Utility Commissions (PUCs) used
an historical year in rate hearings (Williams 1979). This procedure, along with
other factors led to increasing problems with regulatory lag. The solution adopted
was to substitute the lag year with a forward test year, which unfortunately led to
another type of problem: forecast error. Proper financial and operational planning
required that utilities be able to reduce uncertainty in sales forecasts due to
variable fuel prices, volatility in the economy, appliance utilization trends and
behavioral changes due to income changes. Models were developed for both
119
long-term (a year) and short-term (which in this context meant monthly). These
models were mostly econometric models with a combination of behavioral and
engineering models, using regression or time series modeling, or even a
combination of both, called judgmental models (Eslinger and Kirchner 1979).
Note that what econometricians view as engineering models are basically linear
models that contain all the relevant physical regressors (engineers tend to view
these as black box models). Such models have acquired great detail and
sophistication (from an econometrician's view point) and provide detailed
characterization of customers' load patterns at the end-use load level in both
residential as well as commercial buildings. The reader can refer to such EPRI
reports as those by Train et al. (1985) and Limaye and Whitmore (1984).
Methodologies have also been developed to predict peak loads (for example,
Fitzpatrick 1979). Whether available databases and models are still valid under
current-day scenarios and can be used for the purposes of active load management
(the primary focus of this research) needs to be evaluated in the future.
2. Electrical system load forecasting
Electrical system loads are at a regional or distribution-system level, a scale much

larger than an individual building. Usually the forecasts are at a daily level but
have been predicted with tools appropriate for use in single buildings. A review
of pertinent literature is provided in the next section. The tools are of interest for
building load control but the scale is not. These loads are useful in dispatching
generation and managing transmission congestion, duties of an Independent
System Operator.
3. Building electrical load forecasting
The term "building load" can mean one of three things: (i) building thermal loads
(i.e., conduction/radiation loads through the envelope or internal loads, as
described in many textbooks), (ii) heating and cooling secondary system loads
that can be monitored (Haberl et al. 1990), and (iii) the electrical loads in the
building. Our focus is on (iii), but we need also be cognizant of the system
thermal loads since they directly impact certain electrical loads (for example,
setting up the room thermostat reduces the building thermal loads, in turn
reducing the system thermal loads, thereby reducing the electrical load of the fan
(in a VAV system) and the chiller). A review of pertinent literature on (ii) and
(iii) is provided in sections 11.3 and 11.5. Further, our primary intent is to predict
peak load or demand, not on a day-to-day basis, but for relatively few peak days
in the year (say the 5-10 peak days when electric system peaks are hit). Finally,
active load aggregators would be interested in both total electric load as well as
loads disaggregated by end use in order to determine which specific load
reduction measures to initiate.
An important distinction needs to be made in terms of model prediction. Forecasting (or

prediction) is the technique used to predict future values using models that are based
120
upon past and present values (Pindyck and Rubinfeld 1981). There are two types of
forecasts: ex post and ex ante. Both use a certain period for estimating model parameters,
using driving and response variables. In the ex post forecast, also known as an historical
simulation, the forecast period is such that observations of both the driving variables and
the response variable are known with certainty. Thus ex post forecasts can be checked
with existing data and provide a means of evaluating the model during the model
development phase. An ex ante forecast predicts values of the response variable (i.e.,
building energy use) when those of the influential or regressor variables are either:
(1) known with certainty {conditional ex ante forecast), or

(2) not known with certainty {unconditional ex ante forecast).
The unconditional ex ante forecast is more demanding than the conditional ex ante
forecast because the driving variables need also to be predicted into the future (along
with the associated uncertainty that it entails). The studies by Kawashima et al. (1995)
and Seem and Braun (1991) pertain to unconditional ex ante forecasts because only time
series information of past occurrences is included. There are also studies that have
analyzed weather data from numerous locations and proposed hourly sequences of
weather data that can be used for unconditional ex ante forecasts (see for example,
Yoshida and Terai, 1992; Colliver et al. 1995; and Schmitt et al. 2000). Several of the
studies in the literature (for example, Reddy et al. 2002) pertain to conditional ex ante
type of forecasts, although this is not explicitly stated as such. It is obvious that a
conditional ex ante forecast is useful for model development and validation, while the
unconditional ex ante forecast process is the one adopted during subsequent use of the
model.
A load-forecasting tool will rest on the assumption that electricity prices are precisely
known for the forecasting period. In practice, some peak charges are communicated the
morning of their occurrence. Examples include PG&E's Load Management Price Signals
(LMPS), Southern California Edison's 1-6 interruptible rate that is triggered by
curtailment signals from the Independent System Operator (ISO), and load curtailment
signals initiated and sent by electric utilities to members of load-reduction cooperatives.
In this case, the forecast can be updated.
11.2 Short-term forecasts of power-system loads
Short-term load forecasts are an important element of proper control of electrical-power

systems. Both the time scale (typically a day ahead) and the modeling (past loads,
weekday/weekend differences, and weather) are similar to what is involved in forecasting
building loads. This is not surprising, because building loads are a major component of
the total in many power systems. It is worthwhile to briefly summarize the load-
forecasting methodologies, for possible future application to the forecasting of building
loads. Moghram and Rahman (1989) identified the following methods:
• Artificial neural networks (ANN)

• Linear regression
121
• Time-series methods
• General exponential smoothing
• State-space methods
• Knowledge-base methods
Lee et al. (1992) and Ho et al. (1992) both used ANN methods. Lee et al. estimated the
parameters of a nonlinear model using a back-propagation algorithm. Two methods were
used. One method was a day-ahead forecast of 24 hourly loads based on hourly loads in
previous days. Weights for previous days accounted for differences in weekday and
weekend load shapes. The second method took advantage of the autocorrelation of
hourly loads with loads at the same hour on previous days and predicted loads for a single
hour. Both methods predicted day-ahead loads for the Korea Electric Power Company
with relative errors of about 2%. The second method required fewer inputs, trained faster,
and better matched peak loads.
Ho et al. noted that load shapes for the Taiwan power system were well known for
different day types and that load forecasts depended primarily on accurate prediction of
maximum and minimum loads, needed to scale the known load shapes. They proposed to
reduce the training time for their back-propagation algorithm by adjusting the neural-net
momentum, the parameter that determines the emphasis placed on errors in previous
iterations in the back-propagation algorithm, as a function of the magnitude of the
training error. RMS errors for day-ahead load forecasts were less than 1%.
Papalexopoulos and Hesterberg (1990) divided load-forecasting procedures into peak-

load and load-shape models. The former, in limited use, were described as consisting of
a base load and a weather-dependent adder. Load-shape models were characterized as
static or dynamic. Static models consisted of linear combinations of such time-series
functions as sinusoids, exponentials or polynomials, with parameters estimated through
linear regression or exponential smoothing methods. The authors stated that time series
were of limited value because they did not account for weather effects or load
correlations. They favored dynamic load-shape models, which included autoregressive
moving average (ARMA) and state-space models. ARMA models were characterized as
consisting of a deterministic component for the steady-periodic load, depending only on
the time of day, and an ARMA component for random loads. Non-stationary processes
required an autoregressive integrated moving average (ARIMA) model.
The authors described Pacific Gas and Electric's load forecasting algorithm as a weighted
average of ARTMA and multiple-regression peak models combined with an hourly
forecast model that made use of the predicted peaks and the previous day's hourly loads.
They proposed improvements to the model, including modeling of holidays and the effect
of outdoor temperature on heating and cooling loads. They eliminated the ARIMA
model and used only a linear regression. Temperature modeling required a forecast of
temperatures. Errors for peak-load forecasts ranged from -0.17 to -0.89% for the existing
model and 0.09 to -0.51% for the improved model.
122
Charytoniuk et al. (1998) divided load-forecasting methods into parametric methods
(explicit time functions, ARMA models, regressions, polynomials, and Fourier series)
and neural network methods and then proposed a new approach, a nonparametric
regression. The forecast was a conditional expectation of load derived from a probability
density function that related load to time and weather. The model was static and did not
account for significant changes in weather that would produce correlated errors.
Forecasting errors over a three-week period were comparable to those produced by neural
network models, with RMS errors slightly below 4%.
Yang and Huang (1997) noted that neural nets were able to model the nonlinear nature of
load patterns but that a trial-and-error approach is often used to select the structure of the
network and the training parameters. They proposed a fuzzy approach to an ARMA
model with exogenous input, dubbed FARMAX. Heuristics and evolutionary
programming were used to establish the number of input variables, fuzzy partitions of the
input space, and fuzzy membership functions. A local ARMAX model was used for each
set. The FARMAX model produced slightly lower average forecasting errors for Taiwan
Power Company loads than an ARMAX model or a neural net (-2% for each method),
and maximum hourly errors about half those of the other two methods (6.3%, compared
to 13.9 and 10.5%). Automated training was the central advantage of the method.
Ho et al. (1990) forecast short-term loads for the Taiwan Power Company using a
knowledge-based system (KBS). Five years of data, consisting of hourly system loads
and daily high and low temperatures and relative humidities, were used to establish load
shapes for day types proposed by system operators. Peak loads were estimated via a
linear regression against temperature. The expert system was used to assign a day type to
the day for which loads were to be forecasted. The yearly RMS error was 2.9%,
compared to the 4.2% error produced by the Box-Jenkins method, which used an ARIMA
model.
11.3 Forecasts of building electrical loads

11.3.1 Overview
Load-forecasting methods can be classified as follows:
1. Semi-empirical
The aggregator has a rough empirical estimate, for example a kW/ °F metric for
each building, i.e., the rate of demand increase per increase in outdoor dry-bulb
temperature above a certain value. Of course, this is applicable to specific types of
days (for example, for peak summer days, the base could be taken to be 85°F,
which is of course location-specific). Whether such rules of thumb allow end-use
to be also predicted with the required degree of accuracy is uncertain.
123
2. Statistical/adaptive models from historic data
If measured demand data, say hourly data, are available for a year, one could
develop statistical models for different types of days (for example, peak day or
hot days). Techniques such as classical (or cross-sectional) single or multi-variate
models, time series models (EWA or ARBVIA type), or ANN models have been
proposed (for example, Kawashima et al. 1996, Katipamula et al. 1998). A higher
level of sophistication could be to use adaptive time series models (see Seem and
Braun 1991) that allow for self-correction, and thus capture changes in building
operation and occupancy patterns over time. How many buildings have the
disaggregated load data as required by this modeling approach needs to be
evaluated.
3. Simulation-based
The aggregator could also generate predictions based on detailed building

simulation models. The advantage is that disaggregated loads can be accurately
estimated, albeit with more effort since the simulation has to be calibrated first
with actual building load data. Whether this is a viable option needs to be seen.
The forecasting of building electrical loads, which is the primary focus of the load-
forecasting tool, can be broken down in to different aspects:
- Forecasts based on load research data

Whole building electric load models
Disaggregated end use electric models
As previously noted, strategies for demand reduction using the building thermal mass or
cool storage systems have been excluded from this research.
11.3.2 Forecasts Based on Load Research Data
As described in section 11.1, considerable load research data are available. Further,
electric utilities monitor the whole building loads of most of their larger customers at 15-
minute or 30-minute intervals. It would be advantageous for pro-active load aggregators
to make use of this rich source of information, at least at the preliminary level. Akbari et
al. (1988) showed that such data could be used to understand customer patterns as well as
separate the effects of weather-dependent and weather independent effects, both on an
individual customer as well as customer class basis. Another study by Mazzucchi (1987)
described the design and implementation of a large scale monitoring project involving
150 buildings from which key building characteristics and energy use patterns could be
systematically obtained for different building types and classes.
124
11.3.3 Whole building electric load models
Forrester and Wepfer (1984) used multiple linear regression to develop a load prediction
algorithm for the whole-building electricity use of a large commercial building
appropriate for summer use. The algorithm allowed energy and peak use to be predicted
up to 4 hours in advance with an accuracy of 2.5%.
Seem and Braun (1991) reviewed deterministic and stochastic time-series models (which
include AR and ARMA models). The former included polynomial, exponential and
sinusoidal functions. They described a Cerebellar Model Articulation Controller
(CMAC), used by Armstrong et al. (1989) to forecast electricity demand and relying on
an exponentially weighted moving average (EWMA) to update a look-up table used to
map system inputs and outputs. Seem and Braun noted that stochastic time series
methods could be used to model the difference between a time series and a deterministic
model for that time series. Seem and Braun then combined a deterministic and a
stochastic model, and adaptively determined the three autoregression parameters used in
the stochastic model. According to the authors, the algorithm was simple to implement
and its computational and memory requirements were modest. Electrical data gathered
from a grocery store and a restaurant were used to demonstrate the accuracy and
robustness of the algorithm.
11.3.4 Disaggregated end-use electric models
The issue here is whether, and to what accuracy, the major end-use electric loads can be
estimated from whole-building measurements. Three types of development trends have
evolved for common buildings:
1. Econometric modeling such as that by Usoro and Schick (1986), which was
funded by EPRI. The objective was to develop and demonstrate new methods for
estimating load shapes for residential end uses by disaggregating metered data on
whole-house loads. Hourly data over a whole year were obtained from 125
Boston Edison Company customers. A two-level analysis approach was adopted.
At the first level, 60-70 parameters characterizing daily, weekly, seasonal and
weather sensitive patterns of the load were extracted. At the second level, cross-
sectional regressions to measure the influence of household demographics and
appliance ownership were included. This research pertained to residential
customers, but the underlying approach may still be valid to subsequent
development of this tool.
2. Algorithmic approaches such as the SAE (Statistically Adjusted Engineering

method) and the EDA (End-Use Disaggregation Algorithm). EDA (Akbari 1995)
used hourly whole-building data along with certain audit information about the
building and certain physical constraints to produce hourly load profiles for air
conditioning, lighting, fans and pumps, and miscellaneous loads. This approach
125
was applied to two buildings (an office building and a retail establishment) with
an average error of less than 5% during daytime operation.
3. The signal processing approach developed by MIT researchers (described in Luo

et al. 2002 and other papers) where rapid sensing of whole-building electrical use
(waveforms of fundamental and higher harmonics are produced at 120 Hz) along
with sophisticated processing techniques allow individual loads to be detected
with reasonable accuracy.
11.3.4 Building thermal load models
A detailed literature review of building load models has been compiled by Reddy et al.
(1998) under the purview of ASHRAE Research project 1004-RP. Only a brief review is
provided here.
(a) Building simulation models
The building simulation approach relies on adopting a particular engineering simulation

model of energy use in a building and "tuning" or adjusting the inputs of the program so
that simulated output and measured values of building energy use match closely. A
simulation program thus calibrated, could then serve as a more reliable means of
predicting the energy use of the building when operated under different climatic or
different pre-specified operating conditions. One can distinguish between two different
types of engineering simulation models:
1. "detailed", general-purpose, fixed-schematic models such as DOE-2 (Norford et

al. 1989; Bronson et al. 1992; Bou-Saada 1994), and BLAST (Manke et al. 1996);
or
2. "simplified" fixed-schematic HVAC systems models based on the air-side models

developed by ASHRAE TC 4.7 (Knebel 1983) and adopted in slightly different
forms by many workers, for example by Katipamula and Claridge (1993) and Liu
and Claridge (1995). Typically, the building is divided into two zones: an exterior
or perimeter zone and an interior or a core zone. The core zone is assumed to be
insulated from the envelope heat losses/gains, while the solar heat gains,
infiltration heat loss/gain and the conduction gains/losses from the roof are taken
to appear as loads on the external zone only. Given the internal load schedule, the
building description, the type of HVAC system and the climatic parameters, the
HVAC system loads can be estimated for each hour of the day and for as many
days of the year as needed by the simplified systems model. Since there are fewer
parameters to vary, the calibration process is much faster. Therefore, these models
have a significant advantage over the general-purpose models in buildings where
the HVAC systems can be adequately modeled.
126
The detailed-calibrated-simulation-model approach is more tedious and requires expert
knowledge of how the mechanical systems of the building are operated and a certain
proficiency in manipulating the particular building-energy-simulation code. The
detailed-calibrated-simulation approach is typically resorted to under two circumstances:
(1) when the analyst would like to model sub-aggregated electric use from monitored
whole-building monitored energy use, or (2) when the quality or length of the data period
is not adequate to enable proper regression model identification. Both the detailed and the
simplified calibrated model approaches have yet to reach a stage of maturity in
methodology development where they can be used routinely and with confidence by
people other than skilled analysts who developed the models.
(b) Regression models
Model development using the regression approach is generally less demanding in effort
and user-expertise, and yields adequate results. An important aspect in identifying
statistical models of baseline energy use is the choice of the functional form and that of
the independent (or regressor) variables. Extensive studies in the past (for example, see
Fels 1986, Kissock et al. 1998 and Katipamula et al. 1998) have clearly indicated that the
outdoor dry-bulb temperature is the most important regressor variable, especially at
monthly time scales and even at daily time scales.
Classical linear functions are usually not appropriate for describing energy use in many
commercial buildings because of the presence of functional discontinuities, called
"change points." These change-points exist due to the presence of control mechanisms
that include thermostats in residences, HVAC operating and control schedules, and
economizer cycles in commercial buildings (Reddy et al. 1998).
There are in essence two model approaches:
(1) Single-variate (SV) models that have been used to model energy use in commercial
and residential buildings are described in numerous publications (for example, Kissock et
al. 1998, ASHRAE, 1997, Reddy et al. 2002).
(2) Multivariate models. Three basic types of MLR models have been used with some
success to model the hourly variation of the heating and cooling energy use in
commercial buildings:
- Standard multiple linear or change point regression models where the set of
data observations are modeled without retaining the time series nature of the data
(Katipamula et al. 1998),
- Fourier series models which retain the time-series nature of the building-
energy-use data and capture the diurnal and seasonal cycles according to which buildings
are operated (for example, Seem and Braun 1991; Dhar et al. 1998), and
127
- ANN or Artificial Neural Network models (for example, Kreider and Wang
1991, Cohen and Krarti 1997) where automated ANN algorithms are used to model the
time-series trend in a non-linear manner.
How the standard MLR and the ANN fare with respect to the others can be gauged from
the ASHRAE Energy Predictor II (Haberl and Thamilseran 1996) where these and other
approaches were used to model the same data set of monitored building energy use and
then used to predict energy use into a future time period. The availability of monitored
building load data during the future time period allowed the various modeling approaches
to be evaluated against each other in an absolute manner. The standard multivariate
model approach was only marginally less accurate than the ANN models and is much
easier to understand and to use by ASHRAE practitioners.
(c) On-line methods
There are relatively fewer papers that describe on-line models for building thermal loads.
MacArthur et al. (1989) present results of a recursive time series model. Kawashima et
al. (1995) evaluated AREVIA, EWMA, ordinary regression and ANN as unconditional ex
ante forecasts involving predicted weather variables and found ANN to be the most
accurate. Henze et al. (1997) considered various mathematical forms to predict thermal
loads and assess their impact on the performance of a controller for thermal storage
systems. Their load models included an unbiased random walk, a bin predictor model, a
harmonic predictor model, and an autoregressive network predictor model.
Developers of Pacific Northwest National Laboratory's Whole-Building Diagnostician

(Katipamula et al. 1999) switched from a neural net to a set of time series, binned by
temperature, to predict whole-building load.
Daryanian et al. (1994) developed a two-step online procedure for forecasting the day-
ahead hourly cooling load. First, the total load for the next day was estimated on the
basis of the forecasted average outdoor temperature, the total load for the previous day,
and the day type (weekday or weekend). Second, the total load was distributed among
the 24 hours on the basis of historical load-distribution percentages. Regression analyses
showed that outdoor temperature accounted for about 80% of the variation in load and
that the use of three independent variables (temperature, previous load and day type)
produced a correlation coefficient (R2) of 0.95. Forty days of data were used to establish
the hourly load shapes.
11.4 Comparison of models

Hourly building electrical, chilled-water and hot-water loads were forecast by a number
of researchers in a competition sponsored by ASHRAE (Kreider and Haberl 1994).
Contestants were given training data for a four-month period and asked to predict loads
for the following two months. Training data included the building-system thermal loads,
weather data (ambient dry-bulb temperature, absolute humidity, wind speed, and
horizontal insolation) and a date/time stamp. Predictions were assessed on the basis of
128
coefficient of variation and mean bias error. Neural nets, in the hands of experts
typically outside the building-research community and extended beyond conventional
neural-net algorithms, were found to give the best predictions, as indicated by a list of
winners:
1. Bayesian non-linear modeling (MacKay 1994). The model was neural net with an
automatic-relevance-determination model to restrict overfitting of the output to
irrelevant inputs.
2. Feedforward multi-layer perceptron model and a 8-test method for determining

relevant inputs (Ohlsson et al. 1994).
3. An ensemble on neural nets, developed in the petroleum industry to model

relatively small data sets with poorly defined independent parameters and poor
repeatability (Feuston and Thurtell 1994). Several hundred nets with different
typologies and transfer functions were automatically generated and the single net
with the best fit was retained. This method minimized direct labor. Data were
pre-processed with scaling transforms and Principal Component Analysis.
4. Neural nets with two methods, conjugate-gradient and cascade-correlation,

evaluated for training (Stevenson 1994). The number of neurons in the single
hidden layer was determined by trial-and-error for the conjugate-gradient method.
The author noted that predictions would have been improved if the input data had
distinguished weekdays from weekends and holidays.
5. Piecewise linear regression with a steepest-descent method for training (Hjima et

al. 1994). Data preprocessing included data normalization, the removal of
outliers, and the detection and removal of trends, including diurnal periodicity and
differences between weekdays and weekends/holidays.
6. A back-propagation neural net with three-phase annealing to adjust the learning

rate during the training period (Kawashima 1994).
Training is a key factor that affects the performance of neural nets. Back-propagation is
widely used in training the neural nets. It is a mature algorithm and has good
convergence properties (Rumelhard and McClelland 1986). The shortcoming of back-
propagation is that it can take a large number of iterations to converge to the desired
solution. An alternative to back-propagation that has been used in classification is the
probabilistic neural network (PNN) (Specht 1988, 1990), which involves one-pass
learning and can be implemented directly in neural network architecture. A similar one-
pass learning algorithm, called a general regression neural network (GRNN) (Specht
1991), has been used for estimation of continuous variables and was found useful in
learning the dynamics of a plant model for predication or control.
129
11.5 Component Models
11.5.1 Electric power consumed by building air handling units
The theoretical aspects of calculating fan performance are well understood and
documented. Fan capacity and efficiency are calculated from measurements of static
pressure, velocity pressure, air flow rate, fan speed, and power input. Measurement
techniques and calculations are detailed in the ASHRAE, AMCA, and ASME standards
described by Phelan et al. (1997a) as part of ASHRAE research project 827-RP. This
research proposes to base the measurement protocols on the 827-RP recommendations as
needed.
In order to model electricity used by air-handling units, we need to distinguish between

three building air distribution system types and their control (Phelan et al. 1997a):
1. constant air volume (CV) systems;
2. variable air volume (VAV) systems with no fan speed control (i.e., fan operates at
constant speed and flow modulation is achieved by means of dampers); and
3. variable air volume .systems (VAV) with fan speed control (i.e., fan speed is
varied along with damper position to regulate flow; this being more energy
efficient than flow control with inlet vanes.
In CV systems, electricity use is essentially constant during the period during which the
building HVAC system is operated. There may be minor variations in power
consumption as the density of the air varies with changes in the air temperature. Hence a
one-time measurement during the occupied period of the building (and, perhaps, one
during the unoccupied period in case the AHU is shut down) is adequate.
In both cases (b) and (c), electricity use is a function of the building loads, or more
specifically the amount of air supplied to the building. Phelan et al. (1997a) have studied
the predictive ability of linear and quadratic models between electric use and mass flow
rate and concluded that although quadratic models are superior in terms of predicting
energy use, the linear model seems to be the better overall predictor of both energy and
demand (i.e., maximum monthly power consumed by the fan). This is a noteworthy
conclusion given that a third-order polynomial is warranted analytically as well as from
monitored field data presented by previous authors (for example, Englander and Norford
1992 and Lorenzetti and Norford 1993).
11.5.2 Power consumed by pump(s)
(a) General considerations
The theoretical aspects of calculating pump performance are well understood and
documented. Pump capacity and efficiency are calculated from measurements of pump
130
head, flow rate, and pump electrical power input. These calculations and measurement
techniques are detailed in the standards promulgated by ASME and the Hydraulics
Institute, as described by Phelan et al. (1997a). Recommendations on how to perform
pump performance measurements have also been made by Phelan et al. (1997a).
In the same manner as fans, we need to distinguish the operation of circulating pumps in
buildings and in cooling equipment depending on how the pumps are modulated during
part-load operation:
1. constant flow pump with a three way bypass valve to control the amount of heat
transfer in the cooling coils;
2. variable flow pump with a two-way valve to throttle the flow; and
3. variable flow pump with a variable speed drive.
(b) Power consumed by the building pump(s)
In constant flow systems, electricity use is essentially constant during the period in which
the building HVAC system is operated (assuming there are no major pressure variations
in the pumping system). Hence a one-time measurement during the occupied period of
the building (and, perhaps, one during the unoccupied period) is adequate for those cases
where the building pump energy needs to be measured.
However, in both cases (b) and (c), electricity use is a function of the building loads or
more specifically of the fluid flow rate. Phelan et al. (1997a) have studied the predictive
ability of linear and quadratic models between electric use and mass flow rate and
concluded that quadratic models are superior to linear models.
(c) Power consumed by the chiller pump(s)
Normally, there are two separate pumps used by chillers: the condenser pump used to
circulate the water to the cooling tower, and the chilled-water pump that circulates water
through the evaporator. In many cases there are two pumps on the evaporator circuit
(Hartman, 1996): one for the primary circuit, which maintains a constant chilled-water
flow through the chiller, and one for the secondary chilled-water flow to the building,
which is modulated depending on the load.
As discussed by Eppelheimer (1996), controlling head pressure in water-cooled chillers

has long been achieved by varying the condenser flow rate. However, flow rate through
the evaporators normally has not been varied. In recent years, HVAC designers
increasingly have used ways and means of modifying the present day controls in order to
achieve variable evaporator flow and hence better energy efficiency.
131
11.5.3 Power consumed by chiller(s)
The theoretical aspects of calculating chiller performance are well understood and
documented (Gordon and Ng 2000, Reddy and Andersen 2002). There are basically two
types of in-situ models to describe chiller performance: polynomial models and
thermodynamic models.
(a) Polynomial models
This type of model assumes a polynomial function to correlate chiller (or evaporator)
thermal cooling capacity or load and the electrical power consumed by the chiller (or
compressor) with the relevant number of influential physical parameters. For example,
based on the functional form of the DOE-2 building simulation software (LBL 1980)
models for part-load performance of energy equipment and plant, electric use can be
modeled as a tri-quadratic polynomial model (Reddy and Andersen 2002). In this model,
there are 11 model parameters to identify. However, since all of them are unlikely to be
statistically significant, a step-wise regression to the sample data set yields the optimal set
of parameters to retain in a given model.
Braun (1992) used a bi-quadratic model with two regressor variables containing six
empirical coefficients, namely cooling load on the chiller and the difference between the
ambient wet-bulb temperature and the fluid temperature leaving the evaporator (or the
supply temperature to the building). Several other authors (for example, Hydeman 1997)
have also proposed slightly different variants of such polynomial models.
(b) Thermodynamic models
In contrast to polynomial models, which have no physical basis (merely a convenient

statistical one), thermodynamic models are based on fundamental thermodynamic
considerations for a chiller. Such models are preferred because they generally have fewer
model parameters that appear in a functional form with a scientific basis. Furthermore,
their mathematical formulation can be traced to actual physical principles that govern the
performance of a chiller. Hence the model coefficients tend to be more robust, leading to
sounder model predictions. Currently there is only one such thermodynamic model,
which has appeared in two forms, described below.
The complete chiller model proposed by Gordon and Ng (2000) is a simple, analytical,
universal model for the chiller performance based on thermodynamic considerations and
linearization of heat losses. These thermodynamic models were also recommended by
ASHRAE's RP-827 project (Phelan et al. 1997b). The model predicts the dependence of
chiller COP (defined as the ratio of chiller (or evaporator) thermal cooling capacity
divided by the electrical power consumed by the chiller (or compressor)) with certain key
(and easily measurable) parameters such as the fluid (water or refrigerant) return
temperature from the condenser, the fluid temperature leaving the evaporator (or the
132
chilled water supply temperature to the building), and the thermal cooling capacity of the
evaporator. The complete Gordon-Ng model is a three-parameter model that is identified
by multiple linear regression. This model and polynomial models have been studied in
detail by Reddy and Andersen (2002).
11.6 Available software

Private load consultants, load aggregators and utilities have developed and have been
using software tools that pertain to the load-forecasting tool. Although many of these
tools are proprietary, we have been able to identify a report by EPRI (Ried 1987) that
summarizes the capabilities of numerous tools by functionality. Software pertinent to
Tool 5 can be divided into three groups:
1. Load data analysis software that allows processing and analyzing large amounts
of monitored data from a particular facility or buildings. Examples are LDAW
from EPRI (an old version) and a new version by EPRI Solutions, a subsidiary of
EPRI. Another software program is EEM Suite by Silicon Energy. A few also
allow short-term forecasting.
2. Tools that evaluate numerous load management technologies along with their
long-term financial implications for a specific customer. Examples are
LOADSEVI, RELOAD and DSManager.
3. Probabilistic and scenario-based forecasting tools that assess and predict risks
associated with forecasting uncertainties and decomposing sources of errors.
11.7 Illustrative example

Monitored whole-building electric data from a very large commercial building located in
California is used to illustrate the issues involved in forecasting. From the year-long data
available to us, we extracted 10 days of data as shown in Figure 11.1: three days from
June and seven days from October. Because of the weekend, electricity use during two
days (October 6 and 7) was low. We shall attempt to predict hourly electric use during
one day (October 5) using data from two previous days (October 3 and 4).
Figure 11.2 allows us to study the effect of outdoor temperature (weather variable) with
the diurnal schedule of the building. We note the two distinct extremes characterizing the
occupied peak and the unoccupied periods with several observations in-between.
Modeling such a pattern with regression models is awkward since the model will have to
contain not only a temperature regressor but also a dummy variable to account for
occupancy. A more convenient approach is to use a time series model. One of the
simplest time series model is the simple temperature model coupled with an EWMA
average (used by several researchers in the past). The results of such a model fit are
illustrated in Figure 11.3. We note that during the third day the difference between the
model predicted and the monitored values are practically indistinguishable. This
133
illustrates that a simple time series model coupled with a simple regression model may be
adequate for load predictions purposes.
Hottest Periods
(6/20-6/22 and 10/3-10/9)
1000
900
800 -
700 -
600
500
400
300
200 -
™ AmbientTemp(F)x 10
100 WholeBuilding Electric Load (kW)
1 37 73 109 145 181 217 253 289 325 361 397 433 469 505 541 577 613 649 685 721 757 793 829 865 901 937
Time (hours)
Figure 11.1. Time series plots of whole building electric loads during ten of the hottest
days of the year.
Figure 11.2 Scatter plot of whole building electric load versus outdoor temperature
during the hottest days.
134
Using Two days to Forecast the Third
(10/3-10/5)
900
700 - -
600-
500-
400 -
300
100-
MtiHHiiiiiiuimiimiimtitii H))nm)))iiijmiiiini)mi»niiii)iiiiinm>iiiii)ni)ii»»iiiiiinuiiiiiiiiiiiiiimniiiniiiniiiiiiii
1 12 23 34 45 56 67 78 89 100 111122 133 144 155 166 177 188 199 210 221232 243 254 265 276 287
Time (15 min)
Figure 11.3. Monitored versus model- predicted electric use. A simple time series
model was found based on usage of the first two days in order to predict that
during the third.
135
12. COMFORT PENALTY FRAMING (TOOL 6)
12.1 Objective and Scope
The objective of Tool 6, Comfort Penalty Framing, is to propose a procedure that

involves modifying the indoor environment (specifically by allowing the indoor dry-bulb
temperature to increase in a predetermined and controlled manner) so as to reduce
electric demand at the expense of predefined overall occupant comfort. This is slightly
different from the approach described in numerous studies (for example, Braun and
Chaturvedi 2002), which involves pre-cooling the building and making use of the
building thermal heat capacity to shave demand peaks. Such strategies avoid
compromising occupant comfort, while this tool extends this concept further so as to
achieve even higher reductions in demand peaks.
The basic issues involve: (i) formulating the strategy in terms of being able to
mathematically quantify varying degrees of occupant discomfort associated with specific
indoor environment changes given the fuzziness surrounding human comfort modeling;
(ii) assessing whether the technical capability of present day HVAC systems and their
associated controls permit such a strategy to be implemented practically; and (iii)
determining whether the type of building-specific data requirements needed to apply this
tool can be collected for specific circumstances. The procedure may seem drastic and
unacceptable when electricity prices are moderately low, but may be more attractive
when demand charges are very high.
12.2 Literature Review
The quantification of human comfort in general and thermal comfort in particular has
been the subject of several research projects and technical papers, which have resulted in
Standard 55 (ANSI/ASHRAE 1992). Thermal comfort is a cognitive process involving
many inputs influenced by physical, physiological and other processes (ASHRAE 2001).
From the point of view of physical modeling, the most important variables that influence
thermal comfort are: air temperature, air motion, humidity, mean radiant temperature,
metabolic rate and thermal resistance of clothing. The environmental parameters
discussed above are the primary factors used to characterize human comfort. There are
also secondary effects such as nonuniformity of the environment, visual stimuli, age, and
outdoor climate. Despite ongoing work in this general area, this issue still remains
somewhat fluid given: (i) the fact that the parameters such as the insulation value of
clothing and convective heat transfer coefficients that appear in the model need to be
determined experimentally, and (ii) the empiricism and individual differences associated
with climate-chamber experiments on live test subjects.
The most widely accepted approach to physical comfort modeling is the semi-empirical
model based Predicted Mean Vote (PMV) index proposed by Fanger (1972) and
described in ASHRAE (2001). The PMV represents the comfort level of an average
136
occupant of that zone or building. Strictly speaking, it represents an average value of a
large group of people exposed to the actual environment. The PMV is an index ranging
from 3 to - 3 with a value of zero indicating neutral comfort (neither warm nor cool).
ASHRAE (2001) recommends that PMV values be restricted between ± 0.5, but Fanger
(1972) points out that real discomfort is expressed by those voting higher than +2 and
lower than -2.
PMV Perceived
Comfort
3 Hot
2 Warm
1 Slightly Warm
0 Neutral
-1 Slightly Cool
-2 Cool
-3 Cold
The PMV can be calculated using correlations developed by Fanger involving the
metabolic activity level (Mets), clothing constant, mean radiant temperatures of the room
or zone, zone air temperature, zone humidity and clothing surface temperature.
Determination of such quantities requires accurate modeling of the indoor environment,
and is clearly an unsuitable path to follow during practical implementation.
Because the basic model is difficult to use, an empirical correlation for the PMV has been
proposed of the form (ASHRAE 2001):
PMV =a*t+b*Pv+c*
where t and Pv are the indoor temperature and partial pressure respectively, and the
numerical values of the coefficients a*, b* and c* have been tabulated in terms of sex,
and age.
A simpler and more practical approach would perhaps be to perform a census or a survey
with the occupants of each zone, and use these values of PMV for subsequent strategy
implementation in an actual building.
It is the dissatisfied occupants who are likely to complain. And so, rather than use the
PMV directly, Fanger (1972) suggested another index called the PPD (Predicted
Percentage of Dissatisfied occupants). The PPD is directly related to PMV by means of
the following relation:
PPD = 100-95.exp[-(0.03353.PMV 4 + 0.2179.PMV2)] 12.1
In general, the distribution of votes will always show considerable scatter. When the PPD
is plotted versus the mean vote of a large group characterized by the PMV, one typically
137
finds a distribution such as shown in Figure. 12.1. This graph shows that even under
optimal conditions (i.e., a mean vote of zero), at least 5% of the occupants are likely to be
dissatisfied with the thermal comfort. Using the above equation, we have
for PMV=0, PPD=5%

PMV=0.5, PPD=10%
PMV=1.0,PPD=26%
PMV=1.5, PPD=50%
A useful index of acceptability of an environment is the percentage of people dissatisfied

(PPD), defined as people voting outside the range of-0.5 to +0.5. This corresponds to
about 20% of all occupants being uncomfortable in a design indoor atmosphere meant to
satisfy overall human comfort.
%
100 ^ f
"V J
V V
VV J
V
v ->
£ 10 \
T r - —3*
/
PL, • ^
-2 -1
PICT
Figure 12.1. Percentage ofpeople dissatisfied (PPD) as a function of predicted mean

vote (PMV). From ASHRAE (2001).
Despite its fuzziness, the PMV-PPD model is widely used and accepted for design and
field assessment of comfort conditions (ASHRAE 2001). In an office setting, air
temperature and humidity are the two main variables for the physical comfort model.
Gagge et al. (1986) proposed an extension of the PMV method that is more accurate
under sweating conditions. Berglund (1989) found that temperature had an order of
magnitude greater effect on human comfort than did humidity in determining human
comfort. He also found that subjects indicated that equal changes in humidity are more
perceptible at higher humidity levels than at lower humidities. Since load management is
likely to be implemented during hot summer days, high humidity levels and their adverse
impacts on human comfort would be a major issue. An excellent review and results of
138
recent climate chamber test of subjects exposed to high humidity levels has been
provided by Fountain et al. (1999).
12.3 Suggested Strategy and Data Requirements
As stated in Chapter 11, three types of curtailable loads can be considered:
1. Lights and Equipment (On/Off);

2. Motors and HVAC Equipment (Ramped); and
3. Dynamic methods that make use of the thermal capacity of the building mass
and/or cool storage systems.
The portfolio of the aggregator includes several buildings. Within each building, there are
several control panels, and further, there are several circuit loops in each control panel. A
sample electric loop designation chart is shown in Figure 12.2. The electric load at the
circuit loop can be controlled fairly easily if the load type is either on/off or ramped.
The practical implementation of this tool involves first creating a knowledge table like
that shown in Table 12.1. Each circuit loop (see Figure 12.2) is designated by a code and
an index number and the type of load (first three columns of Table 12.1). The load
reduction potential (in kW) of each loop is determined by either measurement or best
guess estimates. Note that on/off type of equipment has one number, while the ramped
type of loads can take on numerical values from 0 to a maximum value. The next step is
to fill the remaining columns in Table 12.1 with comfort-related data.
First, we determine the number of occupants in each loop (which could be one or several
zones depending on how the electrical loops have been installed in the specific building).
Then we assign PPD values to each loop depending on how critical it is to maintain strict
indoor conditions. A certain amount of discretion needs to be exercised by the energy
manager in this regard. Note again, that PPD values corresponding to ramped electrical
loads will have a range from 0 (i.e., no load curtailment) to a maximum value. Finally,
we determine the fraction of uncomfortable occupants (FUO) at each loop:
PPD .N
FUOi =
1
, ' ' 12.2
k
where
k= total number of circuit loops

Nj = number of occupants in each loop i
139
BUILDING CONTROL
SITE LEVEL CIRCUIT LOOPS DESIGNATION
LEVEL PANELS
Bl C1 S1B1C1L1
S1 B2
B3
L1 | •: S2B2C1L1
B1
C1 L2 ->i S2B2C1L2
L3 I •: S2B2C1L3
S2
B2 C2
y*
L2
B3
L1
C3 L2
L3
S3 B2
B3
Figure 12.2. Schematic illustrating the type of disaggregation needed for the Load
Optimization Tool
140
Designation No Load type Load reduction PPD No of FUO Avail.
(0/0,R,D) kW % occupants Status
S1B1C1L1 1 O/O 15 25 15 0.0018 0
S1B1C1L2 2 R 0-25 0-10 40 0-0.0100 1
S1B1C2L1 R 0-22 0-20 45 0-0.0045 1

... D 0-50 0-35 60 0-0.0105 1
Total k - 1,000 - 2,000 1.0 -
Table 12.1. The type of data needed for the Load Optimization Tool
The last column of Table 12.1 simply indicates the availability of electric supply of the
corresponding circuit loop. For example, an availability status=0 may imply that electric
supply cannot be curtailed because of other constraints (conference room or ongoing
experiments) and this loop is not available for demand control at that given time. The
reason of including this additional variable is to allow for flexibility in the load
curtailment strategy, which may vary seasonally, by weekday, or by time of day. Once
such knowledge tables are created, they can be used as a source of data input for Tool 8
(Chapter 14) where the actual load optimization is done (i.e., determining which loads to
curtail such that the stipulated load reduction target is met while minimizing the total
FUO).
12.4 Discussion
ASHRAE's current research programs attempt to link its thermal comfort concepts to
incorporate effects of noise, odor and other elements, and to link comfort with
productivity. Goldman (1999) argues that before attempts are made to extrapolate
ASHRAE's Thermal Comfort Model to include complex interactions with human
behavior, it is more useful to merge it with physiological models of human comfort and
work tolerance. Goldman (1999) also reviews several comfort modeling studies, and
stresses the fact that thermal capacity of the body needs to be included in the model. He
suggests that conditions where the demand D is between 20% and 40% of the body's
capacity C for temperature regulation may cause some human discomfort but would
likely lead to "arousal," i.e., result in greater occupant productivity. However, ratios of
(D/C) > 40% should be avoided since they are associated with decreased performance.
This study seems to suggest that in fact small "thermal shocks" to occupants may
enhance their productivity!
Another accepted design norm can also be questioned. For example, the recommendation
that PMV values between -0.5 and 0.5 be assumed to be the bounds for comfort can be
debated. Goldman (1999) quotes a study by Macintyre that pointed out, based on
experimental data, that the standard deviation for PMV is 1 full scale unit between
exposures (i.e., a PMV of 4 has a 95% confidence range from 2 to 6) and is about 0.8
scale units both within and between subjects for hourly readings throughout a single long
exposure. Hence the normally accepted PMV bounds for comfort are fuzzier than the
design norm. From a practical point of view, one could argue that PMV values of 1
141
(which result in PPD=26%, i.e., 52% of the occupants likely to be uncomfortable) is a
more realistic range to assume. The practical implications of such speculations is that the
normally accepted comfort temperature bounds of 22.2-23.9 °C (72-75 F) can be
extended upwards to 25.6-26.7 °C (78-80 F) during load control periods.
The PMV-PPD approach for modeling thermal comfort is applicable in steady-state only.
In practice, the willingness of an occupant to put up with visually or thermally
uncomfortable conditions depends strongly on their duration and on additional factors,
such as other incentives. For example, occupants might be willing to have office
temperatures rise sharply, but for only one or two hours and for a limited number of days.
During those short periods, workers could schedule meal breaks, seek to work in other
parts of the building (for example, a more massive core area where temperatures rise
more slowly), go outside, or perform less demanding work.
This willingness can be enhanced through financial incentives. For example, County of
Los Angeles jails are permitted to retain energy savings due to responding to an
interruptible rate and use those savings for entirely different budget categories. (Of
course, it might be argued that those detained in a jail are not allowed to express their
thermal comfort vote.)
Thermal sensation complaints are the most common kind of service requests, and the
labor associated with HVAC maintenance can be reduced by 20% if the frequency of
these complaints is reduced (Federspiel 1998). Complaint-based strategies to control
HVAC systems have consequently been suggested. Mac Arthur (1986) proposed using
PMV for controlling thermal conditions in buildings. Federspiel and Asada (1994)
developed a control system based on a modified version of PMV that would respond to
thermal sensation ratings from occupants by adapting to their complaints. Federspiel
(2000) also developed a complaint model to simulate complaint behavior. A recent study
by Martin et al. (2002) present and evaluate different strategies to respond to thermal
sensation complaints. The main cause of thermal comfort complaints has been found to
be unsatisfactory performance of HVAC systems and controls rather than individual
differences in preferred temperature (Federspiel 1998).
Taking full advantage of this approach requires a good model that would combine PPD
with some measure of time or number of occurrences during which occupants will be
subjected to uncomfortable indoor conditions. Such a method could include a weighting
factor that tends to increase the absolute value of PPD for longer periods of time, or a
simple multiplier that expresses willingness of occupants to endure uncomfortable
conditions. The willingness would increase with financial incentive and decrease with
duration.
Hamalainen et al. (2000) describes, in a Finnish context, how to model dynamic

consumption of a coalition of consumers under variable electric prices. Even though this
study pertains to electric heating of residential customers, in contrast to load control in
commercial buildings, it addresses the issue of how to model the consumption strategies
of individual electric buyers so that a certain amount of intentional discomfort is tolerated
142
provided the financial gain is enough of an incentive. This study is the closest in the
literature to the type of strategy proposed in this project. The paper also describes the
physical parameters of sample houses gathered from an experimental field test conducted
in Helsinki during the winter of 1996. It is worth noting that the optimization is based on
a model for human discomfort that merely involves the deviation in actual indoor
temperature from an ideal value, and is not as sophisticated as the approach involving
PMV-PPD advocated here.
12.5 Summary
Any load reduction in the building is likely to compromise occupant comfort over a
period of several hours. Switching off some of the lights and equipment may reduce
visual comfort and needs, while resetting HVAC deck temperatures and/or chiller set
point is likely to impact thermal comfort. The discussion in the previous section indicates
that there is great fuzziness associated with human comfort even under steady state
conditions, and that how humans respond under transient environment as pertinent to
load curtailment measures is largely unknown. Even then, the conclusions drawn from
previous studies suggest that the maximum PMV values that occupants are likely to put
up with for a few hours over a few days in the year could be ± 1, which corresponds to
52% PPD values. This could correspond to indoor temperatures that float up to 25.6-26.7
°C (78-80 F) during load control periods.
The concept of introducing controlled discomfort to occupants in a building goes against

the grain of current thinking and practice. As it is, facility personnel are hard pressed to
maintain occupant comfort and minimize complaints. This is partly due to old or
improper HVAC design, maintenance, control and time. As maintenance, operation and
control of HVAC systems is likely to improve with recent technological developments in
networking and automated fault diagnosis, a strategy such as that outlined here may be
both implementable and acceptable to the facility and occupants (as electric rates increase
and occupants are given alternative compensation for such discomfort).
143
13. INTERACTION OF AGGREGATORS AND
CUSTOMERS WITH POWER MARKETS AND
INDEPENDENT SYSTEM OPERATORS (TOOL 7)
Power marketers are required to purchase power. Mechanisms for doing so are well
established. This chapter will review key points, involving price forecasts, purchases on
the spot market run by power exchanges, and purchases of reserve power as required by
the Independent System Operator. Purchases of power via long-term bilateral contracts
were described in Chapter 10. This review is helpful in the context of pro-active
aggregators and their customers, because it establishes the nature of the price forecasts
needed to facilitate cost-effective load control.
13.1 Price forecasts

13.1.1 Overview
Forecasts of the price of electricity are useful over a wide range of time scales. As noted
in TCA (2000):
• Three-to-five-year forecasts of electricity prices and interruptions allow customers

to assess the financial benefits of decisions to purchase load management systems
or enter into long-term contracts with a service provider to manage their loads.
• Two-to-three-day forecasts of upcoming market price spikes or service

interruptions allow customers to use their load management capability most
effectively. During these few days, a company can adjust production schedules,
arrange for back-up supply, and charge batteries or stockpile fuel so that the load
shedding event has a minimal impact on its operations and profitability.
• Forecasts of prices or interruptions an hour or two ahead allow companies to fine

tune their load management execution - capturing the maximum benefit in the
short-run and avoiding unnecessary disruptions to their daily operations.
The price of electricity at any location at any point in time is a function of the following
factors:
• System load
• Supply curve (quantity vs. price) of available energy
• Transmission congestion
The intersection of the supply curve with the system load is the system-wide price of
electricity in the absence of transmission congestion, resulting from a least-cost, merit-
order dispatch. If congestion makes it infeasible to implement a pure merit-order dispatch
of the system to meet demand, the marginal cost of electric supply will vary from one
144
location to another based on the impact of consumption at each location on flow over the
congested part of the system. If consumption at a given location tends to exacerbate the
congestion, the cost will be higher at that location; if it tends to ameliorate congestion the
cost will be lower.
Each of these elements can be predicted over various time scales, with consideration of
the factors listed in Table 13.1. All of these time scales are of interest for load
management. The multi-year time scale is important for decisions regarding investment
in load reduction technology, the seasonal timescale may be more appropriate for process
scheduling and long-term planning, and the one-to-three-day time scale is used for
optimal implementation of load management strategies.
Time Scale
Cost
1-3 days Seasonal Multi-Year
Component
• Short-term weather
variations
• Business outlook • Demographic trends
• existence of
Load • Climate outlook (El • Market structure,
interruptible or
Nino, etc.) pricing options
price-sensitive
demand
• Fuel Prices
• Maintenance schedules
• Unit availability & • Hydropower availability
• Fuel Prices
maintenance and
Supply • New Entry and
schedule scheduling/operations
Curve Retirement
• Market conditions • Long-term contracts
• Market power
• Market power • Scheduled export/
import
• Market power
• Seasonal variations in
• Transmission line line capacity, etc. • Transmission
and generating unit • Wheeling expansion projects
Congestion outages • Hydropower delivery • Siting of new units,
• Fluctuations in patterns retirements
patterns of demand • Line maintenance • Demographic trends
schedules
Table 13.1. Predictable factors governing electricity prices over various time scales.
13.1.2 Forecasting Sources
13.1.2.1. One-day to three-day forecasts

In some cases, one-day to three-day price forecasts may be available from the ISO. For
example, the New York ISO publishes day-ahead price forecasts on its website. These
145
prices may reflect the ISO's assessment of system conditions, or they may reflect the
buyers' assessment as expressed in the day-ahead market. An example of the NY ISO
day-ahead market price, for the first of 24 hours, is shown in Table 13.2. Capacity,
energy, and ancillary services are scheduled and sold in this market, which closes at 5
a.m. for the following day. The marginal prices over a day (June 3, 2002) for one
regional (Capital) are shown in Figure 13.1.
Location-
based
Time Name of ID marginal price Marginal Cost Marginal Cost
Stamp location number ($/MWHr) Losses ($/MWHr) Congestion ($/MWHr)
6/3/02
0:00 CAPITL 61757 20.87 0.64 0
6/3/02
0:00 CENTRL 61754 19.59 -0.64 0
6/3/02
0:00 DUNWOD 61760 21.83 1.6 0
6/3/02
0:00 GENESE 61753 19.13 -1.1 0
6/3/02
0:00 HQ 61844 20.03 -0.2 0
6/3/02
0:00 HUDVL 61758 21.69 1.46 0
6/3/02
0:00 LONGIL 61762 36.6 1.81 -14.56
6/3/02
0:00 MHKVL 61756 20.09 -0.14 0
6/3/02
0:00 MILLWD 61759 21.59 1.36 0
6/3/02
0:00 N.Y.C. 61761 29.45 1.83 -7.39
6/3/02
0:00 NORTH 61755 19.89 -0.34 0
6/3/02
0:00 NPX 61845 21.2 0.97 0
6/3/02'
0:00 OH 61846 19 -1.23 0
6/3/02
0:00 PJM 61847 18.77 -1.46 0
6/3/02
0:00 WEST 61752 19.07 -1.16 0
Table 13.2. First hour of day-ahead prices as forecast by the New York Independent
System Operator (www, nyiso. com).
146
Figure 13.1. Day-ahead hourly marginal prices for a single location in New York State.
Commercial subscription services also provide short-term price forecasts. One service
offers on its website frequently updated, current-week price and volatility forecasts for
some regions of the country (www.rer.com). Forecasts are made with neural net,
regression, and time-series (ARIMA) models. Another service is based on a proprietary,
short-term forecasting model that uses exogenous load, fuel cost, and unit-availability
forecast data to predict locational market prices (www.tca-us.com). The model accesses a
detailed, proprietary database of generating unit operating data and transmission system .
characteristics.
13.1.2.2. Seasonal forecasts

Forecasting on a seasonal time scale requires detailed knowledge of cyclical variations in
industrial, commercial and residential loads. Another factor that can be important in some
regions is seasonality of hydropower and power exchanges with neighboring regions. The
best sources for these types of information are historical data on energy use and cost.
The other most important, and difficult to predict, factor is variations in fuel prices. The
best source for this information is pricing on forward contracts, which reflect the
market's best estimate of fuel prices. This is notoriously unreliable, however.
13.1.2.3. Multi-year forecasts

Multi-year forecasts are based on an assessment of long-term market dynamics, including
new entry and retirement of generating units, growth in demand and other demographic
trends, and of course fuel prices. Another important factor in the future may be price-
responsive load. This will take the form of both long-term and short-term responses.
Long-term responses are characterized by investment in load-reduction technologies,
such as energy-efficient appliances for consumers and process redesign for industrial
users. These reflect an anticipation of sustained high prices by consumers, but by
bringing loads down they also act to reduce prices in the long term.
Short-term responses are enacted by consumers, generally industrial, who are able to
reschedule their energy use away from peak periods to reduce their overall energy costs.
These responses are also important for multi-year forecasts because their availability is a
147
function of the regulatory environment and consumer investments in load management
hardware and software.
One firm implements long-term forecasts using two tools (www.tca-us.com). The first is
a security-constrained dispatch model designed to compute the hourly locational cost of
electricity based on marginal-cost bidding into a power pool. The second tool analyzes
and forecasts the exercise of market power through strategic bidding and capacity
withholding by generating companies. Both of these take into account the user's
assumptions regarding new entry and retirements, fuel prices, and long-term load growth.
The first tool treats electricity demand as inelastic while the second allows for a
parameterized treatment of elasticity.
Forecasting electricity prices on any time scale requires detailed knowledge of system
characteristics and dynamics. These can be learned from analysis of historical patterns
and research into industry data. It also requires assumptions about future events such as
weather, fuel prices and regulatory actions. Because these assumptions are invariably
problematic, it is essential that any forecast be accompanied by a specific statement of the
underlying assumptions for the forecast to be useful. A forecast of electricity prices is not
a statement of how the future will be, but an analysis of how expected future conditions
will be reflected in electricity prices.
13.2 Bilateral contracts in the forward market
For purposes of this report, bilateral contracts for the purchase of power in the forward
market are considered a contractual issue and not a day-to-day interaction issue, and were
described in Chapter 10.
13.3 Purchases in the spot market
A power exchange is a spot market for electricity. Organizations that need power can
communicate that need and organizations with uncommitted power can specify a price at
which they are willing to sell. Load-serving entities, with a contractual obligation to
provide power to customers, use the spot market to meet their obligations or to sell
purchased power in excess of their needs.
Today, aggregators and their customers do not interact with the spot market. As noted in
Chapter 10, such interaction is a key part of a successful load-control effort. If the
marketer's contract with the aggregator's customers provides incentives for load control,
the power marketer will buy or sell power as usual, but will forecast to the aggregator or
directly to the aggregated customer its need (for purchase) or opportunity (for sale) to do
so. Customers will then need to obtain a price forecast from the ISO or an independent
agency.
148
13.4 Purchases of reserve power as required by the Independent
System Operator
Independent System Operators (ISOs) require power marketers to have available more
power than their customers demand at any time. The ISO establishes the reserve
requirement based on a combination of reserves needed in the event of the loss of the
largest generator or transmission line and a mix of other factors, including uncertainty
surrounding load forecasts and the overall generation mix for the hour. Responsibility
for reserves may be allocated to entities making use of the transmission system in
proportion to the fraction of the system load required by each user. Users - power
marketers and other load-serving entities - may have secured estimated reserves in the
forward market. In general, users may have too little or too much at a given hour and
could then settle among themselves in an ex post clearing market or purchase reserves
from the ISO, which in turn had secured them from generating units. The point of this
discussion is to note that the benefits of load control extend to include reserve power,
because the requirement for reserve power will decrease as load drops. Settlement will
be done after the fact and will need to be included in the financial remuneration
calculations that this report labels as Tool 11.
149
14. LOAD CONTROL VIA GENERATION OF
ALTERNATIVE CONTROL STRATEGIES AND SEARCH
FOR OPTIMUM (TOOL 8)
14.11ntroduction
The results in Chapter 10 identified the potential for load reduction due to several forms
of load control in VAV and CV systems. The results point to suitable forms of load
control in single buildings but fall short in several ways:
• The results, while very useful, were based on steady-state calculations and
ignored such issues as the amount of time required for building temperature to
rise to a higher level after implementation of a load-reduction measure and the
extra cooling energy required to pull space temperatures back down to a normal
level after the load-reduction period.
• The results do not meet the goal of controlling more than one building in a way
that takes advantage of complementary load. Plots of DOE-2 outputs (not
included in this report) suggested that staging chiller use across multiple buildings
would reduce peak charges and reveal the presence of increased chiller loads after
a period of chiller shutdown. However, the results were more illustrative than
exhaustive. At the time this phase of the research was done, the investigators
found no straightforward approach to running many DOE-2 simulations
automatically. The task was not impossible (Huang and Franconi 1999), but
required a heavy investment in manipulating the DOE-2 Building Description
Language (BDL) input files and direct use of the program, as opposed to use of a
third-party version with a convenient user interface. Investigators on this project
were not prepared to mimic that approach.
• DOE-2 generates abundant output. Stripping out what is useful for this research -
energy consumption and temperatures over a peak period - is easily done
manually but again requires a substantial effort to automate.
• The emphasis of the DOE-2 approach is more on a detailed model of the thermal
properties and HVAC systems of a single building and less on an investigation of
optimal control strategies that are centered on temporary reduction of service in
one or more buildings. An important aspect of this work is an exploration of how
building operators could curtail service, in ways that would produce maximum
cost savings and minimum reduction in thermal comfort. DOE-2 is not set up to
allow the user to constrain chiller output to a specified level over a given period
of time.
To meet the goals of the portion of this project that focuses on load control, the
investigators decided to automatically generate a large number of alternative control
150
strategies with a model based on a popular general-purpose engineering mathematics
program (Mathworks 2001). This work is thoroughly documented in Appendix A. This
chapter summarizes the model and results and offers recommendations for future work
using a search engine that automatically interfaces to EnergyPlus.
14.2. Simulation Model
The model is a two-resistor two-capacitor (2R2C) nodal model of a single 100 m2 room
and is simulated over a single day at one-minute time steps. Infiltration is limited to
buoyancy-drive flows. Outside temperature is a pure sinusoid and solar gains are a
sinusoid limited to positive values. Solar and internal gains are split between the air-
temperature node and the structural-mass node, to represent the convective and radiative
portions of these gains. In its initial configuration, the model included a fixed-COP
chiller and a fixed-efficiency fan. The airside system was later extended to include a
VAV system, but only sensible loads were simulation.
The simulation is much less detailed than DOE-2 or EnergyPlus. For the purposes of this
project, it offered two advantages. First, short-term load control measures were readily
programmed. These measures focused on limiting the lighting power and the chiller
power to a specified fraction of the maximum. Second, the simulation was run in a
programming environment that could readily implement a freeware genetic-algorithm
(GA) search engine. The GA search engine made it easy to define an objective function,
which always included energy consumption and optionally included demand and
temperature deviation from set point, and to search over a large number of possible
control strategies for those strategies that minimized the objective function.
14.3. Simulation Protocol and Sample Results
Load control focused on short-term curtailments of lights and cooling that could be easily
implemented and which would reduce operating costs at the expense of thermal comfort.
Luminous comfort was not considered.
14.3.1. Single-Building Optimization
Two sets of simulations were done for a single building: a GA search and an exhaustive
search. First, the GA was used to find the optimal solution, for different forms of the
objective function. Members of the population were defined by the limit on air
conditioning and the hour in which that limit would take place and a similar limit and
hour for lighting. For purposes of the GA, the step size for limits on air conditioning and
lighting was defined as 0.1, ranging from 0 (no service) to 1.0 (no reduction in service).
The load-control period was four hours, noon - 4 p.m.
151
The GA search was performed for the four possible combinations of population size,
three and 10, and number of generations, 2 and ten. In each case, the GA was set up to
remember all past simulations and not repeat a simulation in its search for the optimum.
For example, if the GA had considered a simulation in which the air conditioning was
limited to 0.6 of capacity in hour 2 of the load-control period and the lighting was turned
off completely in hour 3, that simulation would not be repeated if the GA later generated
the same load-control strategy. This approach reduced the number of simulations and
improved performance. A population of ten and a run with ten generations typically
required less than 70 simulations, well below the theoretical maximum of 100.
The GA results were compared with the optimal solution found by an exhaustive search.
The exhaustive search was limited to control steps of 0.2. With this step size and a four-
hour load-control period, the number of possible load-control actions was 576 (six load-
control steps over 4 hours for air conditioning and for lighting). Tables 14.1 and 14.2
show the results of the GA search and how these results compare with the exhaustive
search.
152
run pop gen num Dmd Elec Temp Max A/C A/C Light Light Obj.
sim weight cons dev temp hr cntl hr cntl function
weight weight weight
chk 3 10 29 0 0 0 12 0.7 12 0 78.2080
1 3 2 6 1 0 0 15 0 15 0 84.7307
2 3 10 28 1 0 0 15 0.5 12 0 84.6027
3 10 2 16 1 0 0 15 0 15 0 84.7307
4 10 10 74 1 0 0 13 0.9 12 0 84.5892
5 3 2 6 10 0 0 15 0 15 0 142.1846
6 3 10 27 10 0 0 15 0.1 13 0 141.8331
7 10 2 16 10 0 0 15 0.9 14 0.1 142.0169
8 10 10 62 10 0 0 14 0 15 0 141.8122
9 3 2 6 1 1 0 No No 87.5554
rdn rdn
10 3 10 25 1 0 12 0.6 12 0 85.0090
11 10 2 19 1 0 No 15 0 85.0185
rdn
12 10 10 63 1 0 No 14 0 84.9857
rdn
13 3 2 6 10 0 No No 145.1020
rdn rdn
14 3 10 28 10 0 No 14 0 142.1396
rdn
15- 10 2 16 10 0 13 0.9 12 0 142.4407
16 10 10 63 10 0 12 0.9 14 0 142.1396
17 3 2 6 1 0.1 12 0.6 12 0.4 88.5582
18 3 10 26 1 0.1 12 0.9 14 0 87.5187
19 10 2 19 1 0.1 14 0.6 14 0.1 87.7647
20 10 10 73 1 0.1 14 0.7 14 0 87.5187
21 3 2 6 10 0.1 No 13 0.4 146.0298
rdn
22 3 10 28 10 0.1 14 0.8 14 0 144.6726
23 10 2 20 10 0.1 14 0.9 15 0 145.0054
24 10 10 69 10 0.1 12 0.8 13 0 144.9050
Table 14.1. GA search for optimal control schedule, as a function of population size,
number of generations and weighting Junction. The weighting function now includes a
weight on electricity demand.
153
Runs Demand Energy Cumulative Maximum GA Exhaustive Exhaustive
weight consumption temperature temperature minimum search search
weight weight weight objective minimum maximum
function objective objective
function function
1-4 1 0 0 84.5892 84.5653 88.4935
5-8 10 0 0 141.8122 141.6756 155.9935
9-12 1 1 0 84.9857 84.9768 98.6615
13-16 10 1 0 142.1396 142.0871 166.1615
17-20 1 1 0.1 87.5187 87.5098 102.2813
21-24 10 1 0.1 144.6726 144.6201 169.7823
Table 14.2. Comparison of the best objective function found by GA search with the
minimum and maximum values found via the exhaustive search, for different weights.
14.3.2. Two-Building Optimization
The two-building search took a different form due to the very large number of possible
load-control strategies. Recall that there were 576 combinations for a single building,
using control step sizes of 0.2. For two buildings, the number of combinations is
331,776. This is far too many to consider exhaustively (576 simulations required about 9
hours on a 450 MHz Pin personal computer).
If the two buildings were identical or nearly so, only 576 simulations would be needed to
exhaustively map all the load-control combinations. If the two buildings were not the
same, then 576 simulations would be required for each. The total of 1152 is less than
0.04% of 331,776. The results for simulation #1 for the first building would then be
matched to the objective function from simulations 1-576 for the second building to
determine all possible two-building combinations involving the first load-control measure
for the first building. The same approach would match the objective-function value from
simulation #2 and so on. Simulation results would include energy use for each of the
load-control hours and temperature deviations, and the two-building objective function
would be formed with the coincident demand. The point here is that it is much faster to
form the combinations than to run the thermal simulations needed to generate values for
the objective functions.
It is an open question, not addressed in this research, as to whether the GA could

converge on an optimal solution faster than the time required to form and evaluate all
possible two-building combinations. For this project, two-building load-control was
evaluated by combining limited sets of simulations for a single building. Two sets of
single-building simulations were run, one consisting of three different load-control
combinations and the other of ten. The number of two-building combinations was
therefore nine for the first set and 100 for the second. Objective functions were then
formed with different weights for the two-building hourly energy use and temperature
deviations. Tables 14.3 - 14.6 how the combinations were defined and the results.
Using ten simulations for a single building and 100 two-building combinations produced
154
objective function values that were 0.1 - 4.5% lower. Appendix A analyzes the results in
detail.
Hour Case 1 Case 2 Case 3

AC Light AC Light AC Light
control control control control control control
12 1 1 1 1 1 1
13 1 1 0 0 1 1
14 1 1 1 1 1 1
15 1 1 1 1 0 0
Table 14.3. Control schedules for seeded members of population.
Combination Building 1 case Building 2 case

1 1 1
2 1 2
3 1 3
4 2 1
5 2 2
6 2 3
7 3 1
8 3 2
9 3 3
Table 14.4. Combinations of control actions for two buildings, each of which can be
controlled according to one of three possible schedules.
Run Demand Electricity Cumulative Maximum

weight consumption temperature temperature
weight deviation weight
weight
1 0 0 0
2 1 0 0
3 10 0 0
4 0 0
5 1 0
6 10 0
7 0 0.1
8 1 0.1
9 10 0.1
Table 14.5. Objective-function weights. Each run consisted of an evaluation of the

objective junction for each of the nine combinations shown in Table 14.4.
155
Run Combination
1 2 3 4 5 6 7 8 9
1 161.6315 159.1437 159.1627 159.1437 156.6560 156.6749 159.1627 156.6749 156.6939
2 174.4196 173.0378 171.9302 173.0378 171.6560 170.4960 171.9302 170.4960 169.4614
3 289.5129 298.0844 286.8380 298.0844 306.6560 294.8857 286.8380 294.8857 284.3692
4 162.3226 167.2857 167.0824 167.2857 172.2488 172.0455 167.0824 172.0455 171.8421
5 178.1108 181.1798 179.8499 181.1798 187.2488 185.8655 179.8499 185.8655 184.6096
6 290.2040 306.2264 294.7576 306.2264 322.2488 310.2562 294.7576 310.2562 299.5174
7 167.3886 173.2007 172.9734 173.2007 179.0128 178.7855 172.9734 178.7855 178.5582
8 180.1767 187.0948 185.7409 187.0948 194.0128 192.6066 185.7409 192.6066 191.3257
9 295.2700 312.1414 300.6487 312.1414 329.0128 316.9962 300.6487 316.9962 300.2335
Table 14.6. Variation in objective-function value for each member of the population and
each set of weights for energy and temperature. The minimum and maximum values are
highlighted, the former in bold and the later in bold italics.
14.4. Discussion
This section reviews lessons learned from the simulations in a single building and two
buildings, presents extensions to the basic simulation model designed to extend the
search space to include more load-control options, and makes a case for an alternative
simulation environment for additional work.
14.4.1 Lessons from control of lights and air conditioning in a single

building and in two buildings
Lessons were learned from the use of exhaustive and GA searches, both for a single
building and a pair of buildings:
Single Building
• The peak air-conditioning load increased when air conditioning was curtailed, due
to the pick-up load. If there were a demand charge, the increase in load would
lead to an increase in cost associated with demand. Depending on the magnitude
of demand and energy charges and whether a given day affects demand charges
for a month or year, curtailment of air conditioning may or may not reduce
operating costs.
• While many combinations of lighting and air-conditioning curtailment produced

similarly low values of the objective function (weighted cost of energy and loss of
service), others were poor choices with objective functions that were 15% higher
than the best choices. High objective function values found in this work were
associated with reductions in air conditioning that saved little energy and greatly
increased the thermal-comfort penalty.
156
• The GA was able to find relatively small differences in objective function,
including those due to turning off lights in hour 12 versus hour 15, when the air
conditioning was allowed to stay off for the remainder of the day. It did not
consistently find the absolute minimum when the search space had a shallow
contour, but the extent of suboptimality was trivially small.
• For a problem where there are a large number of near-optimal solutions, the GA
search can be limited to small populations and a small number of generations.
• The same control strategy produced the best performance for all six sets of
weights on demand, energy, and thermal performance: turning off the lights
completely for one hour, which was not subject to a visual-comfort penalty, and
limiting the air conditioning output to 0.6 of maximum value for a single hour. A
second strategy produced the worst performance: turning off the air conditioner
for an hour and maintaining the lights at rated power throughout the load-control
period. This suggests that some amount of off-line testing may establish
heuristics that can usefully supplement or replace time-critical on-line
simulations.
• Reducing lights was always a good idea, because there was no visual-comfort
penalty.
Two Buildings
• Exhaustive searches of two-building combinations were not made, due to the very
large number of possible control combinations. Instead, combinations of three
and 10 choices for each of two buildings were considered.
• Working with a limited set of nine choices produced strategies for two buildings
that were about as good as working with a set of 100 choices. The smaller set
produced objective-function costs that were 0.10-4.45% higher than the costs
found with the larger set, over ranges of weights for energy, demand and
temperature deviation.
• For a given set of weights, the objective function for the worst-case two-building
combination was 2.9-10.2% higher than the best-case combination. The higher
range was associated with large demand charges. Under such charges, it is
important to pay attention to two-building load-control combinations.
• When the objective function included only energy costs, there was essentially no
difference among strategies that involved turning off the air conditioning in the
two buildings during the same or during staggered hours. This was expected,
given the lack of a demand charge. There is no need to use a GA to search for
such a solution. Further, turning off air conditioning was little better than doing
nothing - only a 3% reduction in the objective function, which in this case
represents energy charges. This result was also expected, given the results for a
157
single building. If building owners or operators are willing to curtail cooling for
only a short period (one hour), the savings may be small and require no
consideration of buildings in aggregate.
• Simultaneously turning off lights and air conditioners in both buildings was the
worst option when a demand charge was imposed, due to the coincident pick-up
load following the load-control hour.
• The simulation environment was used to examine two identical buildings. In

principle, the approach should work for dissimilar buildings, including the case
when the peak loads are unequal and are shifted in time, as may be due to
variations in exposure to afternoon solar radiation. Dissimilar buildings were not
tested. It is anticipated that the optimal solution may require the smaller building
to shed load during the time of the peak of the larger building.
14.4.2 Additional control options
It is not necessary to search a relatively small parameter space as exhaustively as was

done in this study. For the two control parameters considered, a limited GA search found
a near-optimal solution with 20-30 simulations, rather than 576. The difficulty lies in
building more options into the simulation and coding the GA chromosome appropriately.
Candidates for extending the set of control options include:
• Indoor temperature set point. The inside temperature could be scheduled in

much the same way as the maximum air-conditioner output. Two numbers in
a chromosome would be sufficient: one to specify the load-control hour and
the other to specify the temperature set point
• Supply-air set point for a VAV system. As noted in Chapter 2, a limited

increase in the set point may lead to an increase in operating costs, because
fan power increases while chiller power remains nearly constant. If the set
point is raised to a level where cooling delivered to occupied spaces is
reduced, electrical power will drop and an optimal operating point can be
found, subject to the magnitude of the thermal-comfort penalty.
The simulation code was extended to include both room-air and supply-air temperature
control. Limited testing was done in both cases; the testing did not include the extensive
investigation that was performed for control of lighting and air-conditioner maximum
power level.
The central problem with more control options is the simulation time required to consider
all possible combinations of load control measures. The goal may appeal: put into play
all possible, easily implemented, short-term load control actions and let a simulation
establish what's best, for a given set of weights on cost and thermal comfort. But the
problem grows in scope so quickly that such an approach is not feasible. Two control
158
actions with 576 choices for each of two buildings produce 331,776 possibilities for two
buildings. Four control actions for a single building, as described in chapter 3, produce
184,320 possibilities. Simulation would require on the order of 200 days for the former
case and 120 days for the latter, with a 450 MHz PIU computer. These times are not
overwhelming if a very small number of off-line simulations were done on a significantly
faster computer to establish heuristics. But it is not practical to simulate such large
numbers of possibilities on a day-ahead basis.
In practice, a building operator would probably not mix such strategies as reducing chiller
output and increasing thermostat set point. Both are intended to reduce cooling load.
The operator would want to know which approach was best, in terms of ease of
implementation, savings, and thermal-comfort penalty. Increasing the room-air
thermostat will provide a known upper limit to room-air temperature and a known
maximum impact on thermal comfort. The thermal dynamics of the building will
determine how long it takes for the temperature to increase. The reduction in air-
conditioning power will be largest when the thermostats are first adjusted. As room
temperatures rise, the cooling load will increase. The reduction in cooling is not
specified or controlled but is the outcome of the change in temperature set point. In
contrast, clipping the output of the air conditioner or chiller yields a known reduction in
cooling energy but an unspecified increase in room temperature. Because the air
conditioning can be controlled at a single point, it should be easier to implement than
changing the temperature set point, at least in buildings that lack digital control at the
zone level. To distinguish the impact of the two control strategies on energy costs and
thermal comfort, it would be reasonable to run relatively small sets of simulations that
would evaluate them individually.
An adjustment in supply-air temperature for a VAV system is more complicated, as

described in Chapter 10. Assume that the supply-air temperature is initially at a level that
minimizes fan and cooling power for a given thermal load and outdoor conditions.
Raising the set point a small amount will increase fan power. The thermal load remains
the same and cooling power will decrease slightly if the chilled-water temperature is also
increased. If the initial set point were indeed optimal, the net effect will be an increase in
plant power. Continue to raise the supply-air temperature set point. Eventually, fan
power will reach a maximum when VAV terminal box dampers reach the wide-open
position. Further increases in set point will cause room-air temperature to rise, reducing
cooling load to the extent that plant power decreases.
14.4.3 Alternative simulation environments
The simulation environment used in this project had the considerable advantage that it
combined estimation of energy use of optimization of control settings. The optimization
routine was easily implemented GA freeware. Energy estimation was made with custom
code. This combination emphasized optimization, appropriate for this study. It was a
better choice than performing a limited number of runs with a standard, publicly available
energy-simulation program and manually searching for optimal combinations of short-
159
term load-control measures. But the energy-estimation package was limited in its
features compared to such programs as EnergyPlus, DOE-2, TRNSYS, and HVACSIM+.
For example:
• The simulation accounted for controller gains and building thermal dynamics.
However, it did not account for controller dynamics (rate-limited actuators) or
HVAC-equipment dynamics (heat-exchanger time constants). Dynamics not
considered could affect peak loads and associated demand charges.
• The VAV system as implemented was a relatively simple representation of an

actual system, lacking an economizer cycle. The computation of air-conditioning
loads was also simplified. Notably, latent loads were not included. Part-load air-
conditioner efficiency and changes in efficiency due to changes in supply-air
temperature were not included.
• The control of air conditioning power, central to most of the simulations of this
chapter, was based on a simplified cooling system with, implicitly, constant
volume. It did not account for changes in fan power associated with reduction in
cooling output.
A simplified energy model appeals because there is hope that it will run faster than a
more capable model that incorporates unneeded features and produces more precision
than may be required. But much effort is required to determine and code the requisite
detail. A better approach for follow-up work, to be supported by ASHRAE or other
organizations, would appear to involve using a recognized energy-simulation code in one
of two ways: coupled directly to an optimization package or coupled to a neural net to
generate simplified results that can then be coupled to an optimizer.
Caldas and Norford (1999), Caldas and Norford (2001a,b), Caldas, Norford and Rocha
(2001), and Caldas (2001) made use of a custom interface between a GA code and DOE-
2. The interface required extensive effort to develop. Changes in the chromosomes as
specified by the GA had to be translated into changes in the DOE-2 input file, a text file
formatted according to the rules of DOE's Building Description Language. This
interface would have required some modification to handle control set points and it was
not available for the ASHRAE project.
GenOpt (GenOpt Manual 2001) is a publicly available interface appropriate for single-
building optimization. It can be readily interfaced with EnergyPlus. EnergyPlus in turn
permits HVAC descriptions that are more similar to TRNSYS in the amount of detail and
flexibility than to DOE-2. This level of detail should support more thorough
investigation of control options than could be done with DOE-2. The combination of
GenOpt and EnergyPlus is recommended for future work.
160
15. POST-CONTRACTUAL CAPABILITIES (TOOLS 9 - 1 1 )
The last set of three tools concerns communication of price and load information,
measurement of load-control action, and financial payments and penalties for load control
or lack thereof.
15.1 Tool 9: Communication Between Aggregator and Customer

15.1.1 Functionality of tool
Tool 9 deals with ways by which load aggregators should set up communication channels
to inform their customers about impending peak days and duration, the price signal, the
possible curtailment measures they could activate, and how much they would save if one
or several of these measures were to be implemented. The customer then needs to
respond to requests, not only with a message indicating compliance or non-compliance,
but also to specify which of the measures he is willing to initiate. For this two-way
communication to work satisfactorily, certain standards and procedures have to be
defined and adhered to; otherwise missed opportunities or oversights could be very costly
to both aggregator and customers. The formulation and development of this tool can
benefit greatly from a recently completed ASHRAE-funded research project specifically
addressing this issue (Kintner-Meyer and Burns 2001a,b).
15.1.2 Literature review
Electric utilities in the regulated market have initiated several load shedding programs in
the past, mainly targeted at curtailing residential air-conditioning. These involve direct
load control (DLC) programs where the utilities, with prior consent of the participating
customers, cycle the air-conditioners either by radio signal or power-line carrier (EPRI
1987). Utilities also have had experience with voluntary price-induced curtailment
programs, such as peak-activated rates, real-time pricing, and time-of-day rates (Orange
and Rockland 1987). Although some of these concepts and hardware are applicable,
major improvements and enhanced capabilities will be needed.
One attractive option is to enhance the capability of the current generation of networked
building automation systems. Although these systems are primarily meant to assist in
proper facility management, which includes energy management, comfort monitoring,
facility operation and service and space planning, it is likely that in the near future they
would include such capabilities as fault detection and diagnosis and predictive
maintenance scheduling, as well as the ability to allow supervisory control essential for
load shedding to be implemented from a central location in the facility. A new field,
called "Enterprise Energy Management" is evolving that involves different levels:
161
1. Baselining and understanding current usage (includes sophisticated reporting
systems that allow the operator to understand and evaluate energy use campus-
wide);
2. Managing enterprise utility usage after the utility contract and reporting process
are in place. This involves real-time monitoring, deciding how best to control load
depending on real-time pricing rates and negotiated rates, and the ability to react
by load shedding;
3. Utility enterprise resource planning which encompasses managing, forecasting

and reacting to dynamic enterprise utility rates.
Building automation systems of the future are likely to include the above capabilities,
while also having the ability to gather weather forecasts and changing utility market rates.
However, all these services are based on the ability to define the energy-related
information services and analysis of data requirements. A recently completed ASHRAE
research project (Kintner-Meyer and Burns 2001a,b) focused on the issues of
communication standardization for utility/customer information services and presented a
comprehensive review of existing architectures and capabilities. Nine information
services (revenue meter reading, quality of service monitoring, real-time price
transmission, load management service, on-site generation supervisory control, energy
efficient monitoring, weather reporting and forecasting services! indoor air quality
monitoring and dynamic demand bidding into a power exchange) were studied. Also
addressed were fundamental characteristics of interoperability in diverse network
environments and three levels of communication, application and scenario
interoperability.
15.1.3 Conclusion/Future
Kintner-Meyer and Burns (2001a,b) have described communication services necessary

for communicating load and price information and have outlined how the
communications tools could be implemented. Some extensions to existing
communication protocols (BACnet, for example, as documented in ASHRAE 1995) are
needed to lay a foundation for tool development.
162
15.2 Tool 10: Monitoring and Verification
15.2.1 Objectives and capabilities
The objective of this tool is primarily to determine, by actual measurement, the amount of
load curtailed by the customer, compared to a previously agreed upon baseline, as a result
of the aggregator initiating a request to curtail load. A second objective of this tool is to
provide the necessary end-use monitored data for both the total-facility baseline load and
the contribution of the important disaggregated end-use loads (needed for Tool 5),
in order that annual adjustments can be made to both these types of uses due to inevitable
changes in building operation, changes in installed loads, and electricity creep. This tool
would thus consist of the necessary sensors, instrumentation, data collection and storage,
and savings verification analysis capabilities.
15.2.2 Literature review
The formulation of this tool can benefit greatly from the maturity acquired by
Measurement and Verification (M&V) efforts in the last 15 years due to skyrocketing
activity of the energy services area. Energy performance contracts often require
verification by actual monitoring of the energy and cost savings resulting from
implementing energy efficiency projects. These M&V efforts are needed because there is
significant uncertainty regarding how well such projects perform relative to their design
predictions. This has been especially true with building systems such as lighting and
secondary and primary HVAC&R equipment and systems whose performance and
control change over time. For these systems, field monitoring is the only way to keep
track of their performance. Several M&V standards and guidelines have been developed
during the last decade, of which IPMVP (2001) and GPC14P (ASHRAE, 2002) are
probably the best and most relevant for this tool development. Other pertinent M&V
programs/documents are US Federal Performance Measurement and Verification
Protocol (USDOE 1996), US EPA Conservation Verification Protocols (USEPA 1996),
the World Bank's monitoring and evaluation guidelines (World Bank 1994), US Initiative
on Joint Implementation (USUI 1996), California's Measurement and Evaluation Tools
(CPUC 1998), LBNL's guideline (Vine and Sathaye 1999) and the North American
Energy Measurement and Verification Protocol (NEMVP 1996).
The IPMVP and the ASHRAE GPC 14P documents merit a brief description:
1. IPMVP (2001, an updated version of the 1997 document) is the result of a

concerted effort between state and federal agencies as well as financial and energy
efficiency experts. It is a consensus document that provides an overview of
current best practice techniques available for verifying results of energy
efficiency, water efficiency, and renewable energy projects. This document
establishes a general framework and terminology to assist buyers and sellers of
M&V services. It is not intended to prescribe contractual terms between buyers
163
and sellers of energy efficiency, although it provides guidance on some of these
issues. Once the contractual issues are decided, this document can help in the
selection of the M&V approach that is most appropriate for the project costs and
anticipated savings, technology-specific requirements, and risk allocation between
buyer and seller.
2. ASHRAE GPC 14P (2002) is a more technical document which complements the
IPMVP document. This guideline provides three methods for measuring savings
from energy conservation retrofits: component isolation, main-meter before-after
measurements and calibrated simulation. Such methods can be used to measure
and verify the payments that should be made to energy service companies,
utilities and others who provide Energy Conservation Retrofit Measures
(ECRMs). It includes using data from specific components, main meters, or
simulated data, and encompasses all forms of energy (electricity, gas, oil, district
heating/cooling, etc.) in residential, commercial and industrial buildings.
Sampling methodologies, metering standards and major industrial process loads
are excluded.
In contrast to the GPC-14P document, which focuses on the relationship of the

measurement to the equipment being verified, the IPMVP document discusses the variety
of monitoring and verification options as they relate to actual contracts for energy
services. Both verification of the conditions for baseline model development and
verification of the quantity of energy savings are addressed. Both documents contain
annexes that discuss issues about different available sensors and data loggers, instrument
accuracy, equipment selection criteria, cost implications, calibration, routine maintenance
and inspection, and data accuracy needs.
There are several research papers, publications and reports that deal with M&V issues.
One useful reference is the special issue by the ASME Journal of Solar Energy
Engineering (Claridge 1998), "Methods of Analysis of Measured Energy Data in
Commercial Buildings." Graphical methods for viewing building energy data, statistical
single-variate models (for example, Kissock et al. 1998) and multivariate models
(Katipamula et al. 1998), Fourier series models, calibrated physical models and ANN are
described, with numerous references listed.
15.2.3 Issues in measurement and analysis
The IPMVP document proposes three monitoring and verification methods or options for
determining energy savings from ECMs either due to retrofits or operational and
maintenance improvements in buildings. The first, Option A, deals with one-time
measurements before and after the ECM. Options B and C, however, deal with
continuous measurements before and after the ECM. Option B typically involves
monitoring specific end-uses for a period of time (for example, weeks to months) before
and after the retrofit, while Option C entails measuring whole-building consumption for a
164
baseline period of at least several months before the retrofit and continuously following
the retrofit.
Although none of the options applies specifically to Tool 10, development of this tool can
be guided by either Option B or C. Because the tool is meant to determine whether and to
what extent the load was reduced by pro-active curtailment, it is necessary to establish a
baseline from which to measure savings. Note the use of the term "measured savings" as
distinguished from audit-estimated savings or other savings measures based primarily on
engineering analysis rather than from the use of measured energy performance data.
Although there is an extensive literature on how to develop baseline models from

statistical considerations (see for example, ASHRAE GPC 14P 2002 or Claridge 1998), it
is unlikely that a statistical approach will apply to Tool 10 because the focus is on
measuring demand savings over a period of a few hours only. The baseline demand
profile for the building or facility is likely to be negotiated by both the energy manager
and the load aggregator based on actual measured load profiles, of say the last five peak
days or some such criterion. Sophisticated statistical (or engineering-based modeling)
techniques are unlikely to be resorted to in the beginning, though their future adoption as
the tool matures is likely. When this does occur, issues such as the uncertainty in the
demand savings (see for example, Reddy et al. 1998 and Reddy and Claridge 2000) will
also acquire relevance.
15.2.4 Summary of salient features of Tool 10
We can group the features and capabilities of this tool in three separate categories as
follows.
(1) Measurement of demand reduction
Direct measurement can be used to determine the amount of load curtailed by the
customer in response to an aggregator-initiated request. This is the most basic capability
and would involve the following tasks:
(a) Install and commission, or recalibrate existing, load meter(s) for whole building
or facility.
(b) Equipment should be in compliance with IEEE standards.
(c) Use measured data from (a) to establish a mutually agreeable baseline for demand
during peak days
(d) Inspect equipment periodically and calibrate as necessary
(e) Process and store all collected data over time. Whenever load curtailment
measures are implemented, corresponding data is used for analysis.
165
(2) Baseline adjustments
The tool should allow adjustments over time to be made to the baseline demand use
during peak days. It is well known that baseline energy use varies appreciably over time
(Claridge 1998). Reasons for this phenomenon include removal or addition of
equipment, addition of such systems as thermal storage that change peak load behavior,
changes in the way the building is operated diurnally and new construction. The tool
should include necessary analysis routines that would use monitored data to make the
necessary changes.
(3) Support for Tool 5: Forecasting Tool
Tool 5 requires that forecast models of the important end-uses be developed (or be
available) so that the aggregator can first determine and then recommend to the customer
the specific curtailment measures to implement. The instrumentation and monitoring
capability should extend to such end use loads. Moreover, even these end-uses would
have to be adjusted over time (as in (2) above), and so all the tasks listed under (1) above
would also apply to each end use.
166
15.3 Tool 11: Financial Remuneration Determination
15.3.1 Functionality
Customers who shift load when requested should be rewarded in accordance with a
contractually specified formula, as noted in Chapter 10. Tool 11 is envisioned as a
formula or collection of formulas. There are a number of issues:
1. Magnitude of cost savings for the aggregate of buildings. The contract with the
power marketer must specify how the cost savings are computed. There are
several options. Most simply, the power marketer could offer a flat rate per kW of
curtailed load. This rate would reflect past experience with spot markets. It
would leave the power marketer with a certain amount of risk that spot prices
would be much higher than anticipated and leave customers with the risk that
sport prices would be lower. More accurately, the power marketer could compute
its costs for spot-market prices relative to bilateral contracts and adjust the total to
its customers.
2. Baselines for each building. Load reductions need to be computed relative to a

baseline. Free riders should be avoided if possible. If, for example, a K-12
school is rewarded for curtailing load during late afternoon system power peaks
but such load curtailment was a routine part of end-of-day equipment scheduling,
the school should not be rewarded. A better strategy is to reward savings relative
to a load baseline established as an average over a small number of recent days
during which there was no load-control activity.
3. Allocation of savings. If savings are not specified on an a priori, fixed payment

per kW basis, there is a need to allocate savings attributed to the group as a whole.
Most simply, customers could benefit according to their average load or their
operating costs. A large customer would benefit more than a small one. More
accurately, customers would benefit according to shifted load, as measured during
times of interest to the power marketer. Payment would be offered on a
percentage basis, with each customer's load shift expressed as a fraction of the
total.
Savings must be calculated on an ex post basis, to account for actual and not
forecasted market prices and the cost that power marketers bear to provide the
reserves required by the Independent System Operators.
15.3.2 Future work
The development of a tool to calculate payments to or from customers is not of itself

difficult. It rests on the availability of load and electricity price information. Load
information in turn depends on appropriate metering. The tool itself is an agreed-upon
formula that would be incorporated in contracts between power marketers, aggregators
and customers.
167
In the past, financial calculations have been implemented at the scale of a local utility or
an entire state (California, for example). Rewards for load control have been offered on a
fixed-price basis, either a one-time payment (as for utility reduction of residential air
conditioning via radio control) or on a fixed price per kW of reduced load. Any lack of
precision in the calculations has been absorbed by the utility or state. Those who
effectively bare the burden of excessive payments to participants are affected at very low
levels, by dint of the large number of non-participants. In a small set of customers served
in aggregate, such imprecisions are likely to make more of a difference. It is therefore
necessary that remunerations be calculated equitably and accurately. Such calculations
are a natural extension, to a smaller scale, of the type of calculations now in effect in
power markets.
168
16. ILLUSTRATIVE CASE STUDY
16.1 Objectives
The intent of this chapter is to demonstrate by means of an illustrative case study the
benefits in multi-building load aggregation. More specifically, we would like to address
the issue of what are the added financial benefits (and under what circumstances) for
each individual building owner to (i) aggregate with others, and (ii) intentionally curtail
his load during peak days in conjunction with the aggregate of buildings.
It is to be pointed out that the scope of this example is not to demonstrate benefits due to
bulk electricity purchase (even though it is a powerful incentive point under the current
deregulated electric scenario), but due to multi-building load aggregation itself. Further,
the case study is meant to compare benefits of curtailing, i.e., reducing load in buildings,
as distinguished from load shifting. Note that load curtailment reduces the magnitude of
the load and does not shift it to, say, off-peak hours. Load shifting makes use of either
cool storage systems or the thermal heat capacity of the building mass itself, measures
that were outside the scope of this project.
16.2 Methodology
The various steps for performing this case study are as follows:
a) Select three different building types and an appropriate geographic location.

b) Choose appropriate values of lights and equipment density, occupancy rates,
schedules, secondary and primary systems.
c) Pick appropriate electric price signals.
d) Use DOE-2 simulation program to perform yearlong simulations. Then, focus the
analysis on a single day of the year with highest electric loads.
e) For this single day, determine the hour-by-hour whole building electric use.
f) Identify a small number of likely load curtailment measures.
g) Repeat DOE-2 simulations and determine the curtailed whole building electric use
for the same peak day of the year.
h) Analyze the results in order to demonstrate financial benefits of multi-building
load aggregation both with and without implementing load curtailment.
16.3 Description of buildings and HVAC&R systems

The geographic location selected was Atlanta, GA, which experiences both hot summers
and cold winters. Table 16.1 identifies prices for energy and demand. The on-peak
demand charge is $10/kW and the off-peak charge is $5/kW with the on-peak hours being
from 11 am till 8 pm.
169
Three different building types were selected: an office building, a hotel and a retail store.
Information pertinent to the three buildings is summarized in Table 16.2. The office
building has 90,000 ft2 and 5 stories, 12 hour/day occupancy, and two screw chillers with
VAV secondary systems. The hotel is 66,000 ft2, with 11 floors, 16 rooms/floor, a lobby,
conference rooms and restaurants, and uses two screw chillers with a four-pipe fan coil
distribution system. The retail store is a 32,000 ft2 building with two stories, and uses
packaged rooftop DX cooling units, one per zone. All three buildings are modeled as
having 10 zones (four perimeter and one core, for both intermediate floors and the top
floor which is coupled to the roof). The building sizes, internal lighting and equipment
densities, ventilation rates, and primary and secondary systems are typical for their
respective building types, and conform to the ASHRAE energy and comfort standards.
The equipment has been sized such that stipulated indoor comfort temperatures are
satisfied more than 99% of the hours in the whole year.
Location Atlanta, GA
Climatic data TMY2
Electric price signals Demand on peak: $10/kW
(on-peak hours: Demand off peak: $5/kW
11 am-8 pm) Energy: $0.10/kWh
Table 16.1 General Inputs.
The diversity factors for all three buildings are shown in Figure 16.1. Note that the hotel
peak occurs in the evening while those for the other two buildings occur during normal
working hours.
DOE-2 runs were performed for each of these buildings under the baseline conditions
described in Table 16.2. It was found that July 3 had the highest electric peaks for all
three buildings, and so this day was selected for subsequent analysis. The outdoor
temperature and the horizontal solar radiation are plotted in Figure 16.2 for this day. Also
shown in Figure 16.2 is the demand rate identified in Table 16.1.
The hourly electric load profiles for all three buildings during July 3 rd are shown in
Figure 16.3. While the peak for the Office is around 530 kW, that for the Hotel is about
240 kW (at 8 pm) and that for the Retail Store around 320 kW. The individual building
hourly internal loads, consisting of lights and equipment, are also shown for comparative
purposes in Figurel6.3.
170
BIdg Type Office Hotel Retail Store
General
Total size (ft2) 90,000 66,000 32,000
No. of stories 5 11 2
Dimensions (plan) 220'x 90' 120' x 50' 200' x 80'
No. of zones 10 10 10
Other info Type A 16 rooms/floor

+lobby +conf.
+restaurant
BIdg Roof R-20 R-20 R-20
envelope Wall/Windows 0.479 0.479 No windows
UvaIue(Btu/F-ft2)
Internal Schedule 12 hr/day 24 hr/day 14 hr/day
(see Fig. 16.1)
Lighting 1.5 2.0 2.0
(W/ft2)
Equipment 2.0 0.25 0.50
(W/ft2)
Occupant density 200 ft2/person 200 ftVperson 300 ftVperson
Ventilation Occupied 20 cfm/person 40 cfm/person 10 cfm/person

Unoccupied 0 cfm/person 40 cfm/person 0 cfm/person
Thermo Occupied (°F) 75 72 74
setting* Unoccupied (°F) 85 72 85
Primary Screw chillers Screw chillers Packaged
equipment** (2) (2) rooftop
Size 300T 240 T One per zone
Secondary - VAV 4 pipe fan coil DX cooling
Heating Gas boiler (space Hot water Yes Yes

heating) reheat (furnace)
Gas water heater Yes Yes Yes
(DHW)
* Dead band of 3 °F with PI control

**Secondary and primary systems sized such that indoor conditions were satisfied 99 -
100% of the time on an annual basis
Table 16.2. Description of buildings and systems simulated.
171
Schedules for Lights and Equipment
100
- - • - - Office-LE-Dtversity
90
— — • Retail-LE-Diversity
r ~ „__.,.
;j }TV] T
60
"•j H---\-
40
30
10
12 20
Hour
Figure 16.1 Diversity factors for lighting and equipment for the three buildings.
— • — To(F)
- • - Solarx0.1 (Btu/hr-ft2)
—-•—Demand ($/kW)
12
Hour (July 3rd)
Figure 16.2. Hourly outdoor temperature, horizontal solar radiation and demand rate for
the peak day used for analysis (July 3).
172
Office Building- Baseline
600
>. 3 0 0 -
200 -
(a) Office Building

Hotel- Baseline
• - Internal Total Elec (kW)

Total Bldg Electric (kW)
400
200
•-•-•-•-*
12
Hour (July 3rd)
(b) Hotel
173
Retail Store- Baseline
600
- • - Internal Total Bee (kW)
• Total Bldg Electric (kW)
•1 a •
(c) Retail Store
Figure 16.3 Hourly baseline profiles of total building electric use and the internal
electric loads (lights and equipment) during the peak day
16.4 Load curtailment measures selected

As noted in section 16.1, the load curtailment measures considered in this study are those
that will result in load reduction, not load shifting. Four different measures were
selected, as summarized in Table 16.3. These involve reducing the lighting and
equipment electric density levels, changing the thermostat settings, and changing the
ventilation rates during the occupied hours. The new values selected are lower than the
ASHRAE recommended standards (and in that sense indoor human comfort may be said
to be compromised), but special care has been taken to keep them reasonable and not
select extreme values so as to make this case study more compelling in its conclusions. In
that sense, the curtailment measures selected are deemed to be realistic ones.
174
Office Hotel Retail Store
Baseline Load Baseline Load Baseline Load
Curtailment Curtailment Curtailment
Lighting 1.5 1.0 2.0 1.0 2.0 1.0
(W/ft2)
Equipment 2.0 1.0 0.25 0.25 0.50 0.50
(W/ft2)
Thermostat 75 78 72 76 74 77
setting (°F)*
Ventilation* 20 10 40 30 10 7.5
(cfm/occup.)
* during occupied period
Table 16.3 Load curtailment measures selected.
16.5 Analysis of DOE-2 simulation results

The DOE-2 simulation program was run for the whole year for each building separately,
both without (baseline case) and with load curtailment, measures. The hourly total
building electric use for baseline operation and when load curtailment measures are
implemented are shown in Figures 16.4 and 16.5 for individual buildings and for the
aggregate respectively for the peak day of July 3. The pertinent numerical results for July
3 are summarized in Table 16.4. We note that the on-peak value for the Office building
was 535 kW and occurred at 4 pm. Those for the Hotel and the Retail occurred at 8 pm
and 6 pm respectively, with the aggregate of all three buildings occurring at 4 pm. The
"Individual Sum" column relates to the case when the three buildings are not aggregated,
i.e., if they are billed individually. Hence, 1,090.7 kW is the sum of the individual electric
peaks of the three buildings. While the sum of the individual buildings resulted in a
demand charge of $16,191, that of the aggregate was 8.4% less, because of the non-
coincidence of the peaks in the three buildings. It is interesting to point out that this value
is consistent with the three actual case studies documented in Phase I. The second half of
Table 16.4 summarizes results when load curtailment measures are implemented in the
three buildings. We note from Table 16.4 that the cost savings due to aggregation under
load curtailment is 6.1%. Finally, the cost reduction under load curtailment and
aggregation as compared to baseline operation of the three buildings without load
aggregation is 36.3%.
The above analysis was repeated for the same demand signal but with the on-peak period
being only from 1 pm till 6 pm as against 11 am - 8 pm as assumed previously. The end
results were essentially similar to those shown in Table 16.4, thus suggesting that on-
peak duration is not a major factor in multi-building load aggregation.
175
It will be noted that the above numbers are for demand charges only, and do not include
energy charges. Since there is no energy benefit when multi-buildings are aggregated, we
have intentionally chosen not to include energy charges in the above analysis.
176
Office Building
600
500
400
>. 300
200-
100
• -•-•-•.• • .•-<>
12 20
Hour (July 3rd)
(a) Office Building

Hotel
600
- • - Curtailed
Wt •" Baseline
500
400
3
>. 300
200
100
12 16
Hour (July 3rd)
(b) Hotel
111
Retail Store
600
500
400
>. 300
200
12
Hour (July 3rd)
(c) Retail Store
Figure 16.4. Hourly total building electric use during the peak day (July 3) for baseline
operation and when load curtailment measures are implemented.
178
Aggregate of all Three Buildings
0 4 8 12 16 20 24
Hour (July 3rd)
Figure 16.5 Hourly aggregated electric use for all three buildings during the peak day
(July 3) for baseline operation and when load curtailment measures are implemented.
Office Hotel Retail Individual Aggregate

Sum
Baseline On-peak 535 234.5 321.2 1,090.7 988
(kW) (4 pm) (8 pm) (6 pm) (4 pm)
Off-peak 522.6 211 323.3 1,056.9 990
(kW) (6 pm)
Demand $7,963 $3,400 $4,828 $16,191 $ 14,829
charge ($)
Cost Savings 8.4%
Fraction
With load On-peak 314 154.4 279.6 748 699
curtailment (kW) (4 pm) (8 pm) (2 pm) (4 pm)
Off-peak 310 140.9 249.3 700.2 664
(kW) (9 am)
Demand $4,690 $2,248 $4,042 $10,980 $10,310
charge ($)
Cost Saving 6.1%
Fraction
Demand Saving Fraction due to aggregation and load curtailment measures 36.3%
Table 16.4 Summary of the simulation results.
179
July 1999 Peak Days
1,000
-July5th (Monday)
• July 6th (Tuesday)
• July 19th (Monday)
•July 23rd (Friday)
- July 15th (Thursday)
800
600
I-
200
L J
-• T -- T --- l r --~- r -T '
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
-200
Time
Figure 16.6. Spot electric price signals for 5 days in July 1995 for the PJM electric grid.
16.6 Concluding remarks

Electric price signals under a deregulated market may be one of two kinds: RTP or on-
peak/off-peak differential demand rates (which may be seasonal). The simplest type of
RTP signal is one where the demand charges are combined with energy charges and a
combined price is quoted on a $/kWh basis. Such plots are shown in Figure 16.6 for the
PJM grid for 5 days in July. Notice the large diurnal shifts between the five days. Of
more relevance to us is that, under such a pricing signal, there is no incentive to aggregate
load across buildings provided we overlook discounts due to bulk purchases and assume
that no load shifting measures are being implemented in the buildings. The real benefit in
load aggregation is in demand reduction, i.e., the total aggregated demand across several
buildings will be less (at most, it can be equal) than the sum of the individual building
peaks. This is how the individual building owners would all benefit from aggregation.
A noteworthy observation from this DOE-2 analysis is that the results are consistent with
those found from the three real-life case studies described in Phase 1 of this study. We
find that multi-building load aggregation (with no load shifting) results in 6 - 9%
reduction in demand costs as compared to when each building is operated and billed
separately.
180
17. CONCLUSION
The first phase of the research examined the current state of load aggregation. Specific
tasks involved (1) an extended background and historical perspective on electric utility
aggregation, (2) a description of the various candidate sites where load aggregation
programs have been implemented and clearly recorded, (3) a detailed description of each
of the three case-study sites where load aggregation and subsequent energy management
strategies were performed, and (4) a summary of the lessons learned from the above case
studies. There was little evidence of active load control within the examined aggregates
of buildings. Instead, aggregation was driven by a desire to obtain lower rates through
the magnitude of the load offered to power marketers and through the benefits of
metering that measured coincident rather than individual-building peaks. However,
incentives for load control within aggregates of buildings are being developed as the
recently deregulated electric power market progresses to maturity.
The second phase of the research identified a series of tools that are needed to support the
control of loads in aggregates of buildings. Many of the tools are also required to
effectively control loads and costs in single buildings. For example, owners and
operators of a single building, not participating in a power-aggregation consortium, have
a need to understand their load-reduction potential, to enter into a contract with a power
marketer, to compute thermal comfort penalties associated with short-term load-reduction
strategies, to receive appropriate price signals, to optimally control loads, to record load
reductions, and to be appropriately compensated. Several tools are associated entirely or
in part with aggregates of buildings, including the pre-screening and portfolio
optimization tools and those aspects of the contractual, optimization, and remuneration
tools that must account for multiple buildings.
At a more basic level, several of the tools are needed to provide power to customers who
have no ability to shed load. Even these customers require a power contract, the terms of
which may depend on their load shape and on the load shape a power marketer must
satisfy via purchases from generators. Contracts and a minimal amount of measurement
are fundamental to electricity service. Remuneration is built into the contract negotiation
process, in the sense that a power marketer competing aggressively for business may
want to share cost reductions associated with serving favorable load shapes.
The more technical tools, for assessment of load-reduction potential, calculation of

thermal-comfort penalties, and optimization of load, represent a substantial step forward
but would also benefit from more work. They are linked by a common premise: that it is
reasonable to consider short-term, easily implemented load control measures and trade
their benefit against a short-term reduction in thermal comfort. Assessment of load-
reduction potential is most accurately done on a building-specific basis with a dynamic
building-energy simulation program. The calculation effort is likely to be higher than the
market will bear. Instead, two approaches have been recently developed. Dynamic
simulations have been used to estimate load reductions in generic buildings in a variety of
California climates (Haves 2002). This project took a different approach, using a
181
spreadsheet with building-specific inputs. The spreadsheet was employed to assess load
reductions for constant-volume and VAV systems under a range of internal loads, for an
increase in zone temperature, a reduction in lighting loads, and an increase in supply-air
temperature. The spreadsheet is available from the Principal Investigators.
Thermal comfort has been thoroughly studied. What is proposed here is a method for
estimating thermal comfort at the most local level (an individual office or thermal zone -
wherever zone temperature data are available) and then forming an occupancy-weighted
estimate of thermal comfort within an entire building. Such a weighted average would
permit an assessment of the thermal-comfort penalty associated with zone-specific load
control measures, including increases in zone temperature. This approach has not been
implemented in buildings. It would be worthwhile for ASHRAE to sponsor research to
measure changes in thermal comfort in response to short-term load control measures. Of
further note is that this type of control introduces a dynamic aspect to thermal comfort,
because occupants will be exposed to increased temperatures and humidities for
relatively short periods (a few hours). While the body may physiologically respond
rapidly, occupants who are informed of the reason for the load control and its anticipated
duration may adapt psychologically and experience minimal discomfort.
The optimization tool developed for this study is, to the principal investigators'
knowledge, the first attempt to look at load control across multiple buildings. The tool
uses a simplified dynamic simulation and a genetic algorithm. At a single-building level,
the tool can be used to search for an optimal control strategy, choosing among a set of
short-term control measures. It is superior to a set of guidelines because it is building
specific and is superior to steady-state methods because it accounts for thermal dynamics.
It has two major drawbacks. First, it substitutes a brute-force combinatorial approach for
a certain amount of wisdom. All possible strategies, such as increases in zone-air
temperature or supply-air temperature and reduction of chiller capacity, are considered
equally. The number of possible control options is large, given a range of set point
adjustments for each measure and a range of hours of implementation. However, in
practice all measures are designed to allow zone temperatures to rise and fan and/or
cooling power to drop. It is likely that the measures would not be implemented in
combination in practice. Second, "the building-energy simulation is not as comprehensive
as EnergyPlus. It is recommended that EnergyPlus be used with available search
engines.
At the level of combinations of buildings, the search for an optimal control strategy must
recognize that the number of control options can easily grow to the point where
computation takes on the order of weeks rather than minutes. This project made use of a
simplified search for the optimum strategy for two buildings, taking a randomly
generated set of options for a single building and exhaustively evaluating the small
number of two-building combinations.
Load control is not currently part of most load-aggregation programs, which are focused
on buying power and obtaining the best price for electricity. There have been and are
exceptions. Local utility companies have in the past sponsored so-called cooperative
182
programs, in which participants are contractually obligated to shed a specified amount of
load as a group in response to utility notification. California has recently rewarded
businesses that have shed load when requested. While buildings have acted on their own,
the program is intended to benefit the state as a whole, which might be considered a
large-scale aggregation of buildings. Exceptions aside, it is worth asking why load
control is built into load aggregation. One answer is that neither simulations nor field
experiments have been brought to bear, to the extent that savings from load control are
documented and assessed. Another answer is that the contractual, measurement, and
bookkeeping tools are not in place. These tools, which in effect sandwich the technical
tools (or, put another way, form the front and back ends of a load-control program), can
be envisioned as junior versions of tools that have recently been developed for use in
power markets. The scale is different - much smaller. Potential developers of these
tools, which could be aggregators or their consultants, need some assurance that there is a
market for load-control programs.
As power markets, with their financial settlement instruments, mature, it will be easier to
push the accounting tools to lower levels. More widespread use of more capable
electricity metering and communications technologies will provide a more extensive
infrastructure for load control. Publicly available optimization engines coupled to
building-energy simulation programs will also extend the infrastructure for load control.
To accelerate the implementation of effective multi-building load control the following

research and development is recommended:
• Tests of short-term load-control measures in commercial buildings. Ease of

implementation, load reduction and increase in temperature at a zonal level should
be measured. Pick-up loads should also be recorded, for purposes of assessing
how to stagger load-reduction across multiples of buildings.
• Load-reduction potential tools should be developed to the point where aggregators

can use them with confidence. This is now important in the California Power
Authority's Demand Reserves Partnership program.
• Optimal-control tools should be further developed for single- and multi-building

short-term load-control measures. These tools should account for thermal- and
visual-comfort penalties.
• A financial remuneration methodology should be established to account for the

contribution of an individual building to the load reduction of the aggregate.
183
REFERENCES
Akbari, H., 1995. "Validation of an Algorithm to Disaggregate Whole-Building Hourly
Electrical Load into End Uses," Energy 20(12): 1291-1301.
Andersen, I. And M.J. Brandemuehl. 1992. "Heat Storage in Building Thermal Mass: A
Parametric Study," ASHRAE Trans. 98(1).
ANSI/ASHRAE Standard 55. 1992. Thermal Environmental Conditions for Human

Occupancy, American Society of Heating, Refrigerating and Air Conditioning Engineers,
Atlanta, GA.
ASHRAE. 1995. "ANSI/ASHRAE Standard 135-1995. BACnet- A data communication

protocol for building automation and control networks," ASHRAE, Atlanta.
ASHRAE. 1997. Handbook of Fundamentals, American Society of Heating,

Refrigeration and Air-Conditioning Engineers, Atlanta, GA.
ASHRAE. 2001. Handbook of Fundamentals, American Society of Heating,

Refrigerating and Air Conditioning Engineers, Atlanta, GA.
ASHRAE. 2002. Proposed Guideline GPC 14P, Measuring Energy and Demand
Savings, under public review, ASHRAE, Atlanta.
Berglund, L., 1989. "Comfort criteria in a low humidity environment," Final report,
Electric Power Research Institute, Palo Alto, CA.
BLAST 2002. http://www.bso.uiuc.edu/BLAST/
Bou-Saada, T., 1994. "An Improved Procedure for Developing a Calibrated Hourly
Simulation Model of an Electrically Heated and Cooled Commercial Building", Master
of Science thesis, Mechanical Engineering Department, Texas A&M University, College
Station, TX.
Braun, J.E., 1990 "Reducing Energy Costs and Peak Electrical Demand Through Optimal
Control of Building Thermal Storage," ASHRAE Trans. 96(2), paper # SL-90-16-2.
Braun, J.E., 1992. "A Comparison of Chiller-Priority, Storage-Priority, and Optimal

Control of an Ice-Storage System", ASHRAE Trans. 98(1):893-901.
Braun, J.E. and N. Chaturvedi, 2002. "An Inverse Gray-Box Model for Transient
Building Load Prediction," Int. J. ofHVAC&R Research 8(1): 73-99.
184
Bronson, D., S. Hinchey, J. Haberl, J., D. O'Neal, and D. Claridge, D. 1992. "A
Procedure for Calibrating the DOE-2 Simulation Program to Non-Weather Dependent
Measured Loads," ASHRAE Trans. 98(1), paper # AN-92-1-5.
Caldas, L. G. and L. K. Norford. 1999. "A Genetic Algorithm Tool for Design
Optimization." ACADIA '99. Proceedings of the 1999 Conference of the Association for
Computer-Aided Design in Architecture, Salt Lake City.
Caldas, L. and L. K. Norford. 2001a. "Architectural Constraints in a Generative Design

System: Interpreting Energy Consumption Levels." Proceedings of Building Simulation
2001, International Building Performance Simulation Association.
Caldas, L. G. and L. K. Norford. 2001b. "A Design Optimization Tool Based on a

Genetic Algorithm." Automation in Construction 11(2) 173-184.
Caldas, L., L. Norford, and J. Rocha. 2001. "An Evolutionary Model for Sustainable
Design." Proceedings of Passive and Low Energy Architecture (PLEA) 2001.
Caldas, L.G. 2001. "An Evolution-Based Generative Design System: Improving the
Environmental Performance of Buildings." Ph.D. thesis, Building Technology Program,
Department of Architecture, Massachusetts Institute of Technology, Cambridge, MA.
Charytoniuk, W., M.S. Chen and P. Van Olinda. 1998. "Nonparametric Regression
Based Short-Term Load Forecasting." IEEE Transactions on Power Systems, 13(3):725-
730. August.
Claridge, D.E., (ed) 1998. Special Issue on Conservation and Solar Buildings, ASME
Journal of Solar Energy Engr. 120(3). August.
Clark, D.R. 1985. "HVACSIM+ Building Systems and Equipment Simulation Program
Reference Manual." NBSIR 84-2996. National Institute of Standards and Technology,
Gaithersburg, MD.
Clark, D.R. and W.B. May, Jr. 1985. "HVACSIM+ Building Systems and Equipment
Simulation Program - Users Guide." NBSIR 85-3243. National Institute of Standards
and Technology, Gaithersburg, MD.
Colliver, D.G., H. Zhang, R.S. Gates and K.T. Priddy, 1995. "Determination of the 1%,
2.5% and 5% Occurrences of Extreme Dew-Point Temperatures and mean Coincident
Dry-Bulb Temperatures," ASHRAE Trans. 101(2) paper #3904 (RP-754).
Cohen, D.A. and M. Krarti. 1997. "Neural Network Modeling of Measured Data to
Predict Building Energy System Retrofit Savings," Proceedings of the 1997ASME
International Solar Energy Conference, p.27, Washington, D.C., April.
185
CPUC. 1998. "Protocols and Procedures for the Verification of Costs, Benefits, and
Shareholder Earnings from Demand-Side Management Programs," California Public
Utilities Commission Report, San Francisco, CA, March.
Dhar, A. T.A. Reddy and D.E. Claridge. 1998. "Modeling Hourly Energy Use in
Commercial Buildings with Fourier Series Functional Forms," ASME Journal of Solar
Energy Eng. 120(3):217-223.
Daryanian, B., L.K. Norford and R.D. Tabors. 1994. "Automatic Control of Thermal
Electric Storage (Cool) Under Real-Time Pricing." New York State Energy Research
and Development Authority report 94-13.
Decision Engineering, 1999. Crystal Ball Software, version 4.0.
DOE2 2002. simulationresearch.lbl.gov/
EnergyPlus 2002. simulationresearch.lbl.gov, www.energyplus.gov.
Englander, S.L. and L.K. Norford. 1992. "Saving Fan Energy in VAV Systems- Part 1:
Analysis of a Variable-Speed-Drive Retrofit," ASHRAE Trans. 98(1):3-18.
Eppelheimer, D.M., 1996. "Variable Flow- The Quest for System Efficiency," ASHRAE
Trans. 102(2):673-678.
EPRI. 1985. "Hourly Electric Load Model (HELM), vols. 1-3, Electric Power Research
Institute, publication EA-3698, Paolo Alto, CA, January.
EPRI. 1987. "Residential Load Control and Metering Equipment: Costs and
capabilities", EPRI Report EM-5392, prepared by Electrotek Concepts Inc.,
Mountain View, CA.
EPRI. 2000. "PowerShape, Version 2.0," Electric Power Research Institute, Palo Alto,
CA, March.
Eslinger, R.D. and R.L. Kirchner, 1979. "Short term Sales Forecasting Using Time Series
Analysis," Symposium Proceedings "How Electric Utilities Forecast," EPRI Report EA-
1035-SR, Special Study Project WS-77-46, March.
Evans, J. and D. Olson, 2000. Statistics, Data Analysis and Decision Modeling, Prentice
Hall, Upper Saddle River, NJ.
Fanger, P.O. 1972. Thermal Comfort. McGraw Hill, New York.
Federspiel, C.C., and H. Asada, 1994. "User-adaptable comfort control for HVAC
systems," Journal of Dynamic Systems, Measurement and Control 116(3): 474-486.
186
Federspiel, C.C., 1998. "Statistical Analysis of Unsolicited Thermal Sensation
Complaints in Commercial Buildings," ASHRAE Trans. 104(l):912-923.
Federspiel, C.C., 2000. "Predicting the Frequency of Hot and Cold Complaints in
Commercial Buildings," Int. J. of HVAC&R Research 6(4):912-923.
Fels, M. (Ed.), 1986. "Special Issue Devoted to Measuring Energy Savings, The
Princeton Scorekeeping Method (PRISM)," Energy and Buildings 9(1-2).
Feuston, B.P. and J.H. Thurtell. 1994. "Generalized Nonlinear Regression with
Ensemble of Neural Nets: The Great Energy Predictor Shootout," ASHRAE Trans.
100(2): 1075-80.
Fitzpatrick, G.L., 1979. "Peak Load Forecasting Methodology," Symposium Proceedings

"How Electric Utilities Forecast", EPRI Report EA-1035-SR, Special Study Project WS-
77-46, March
Fountain, M.E., E. Arens, T. Xu, F.S. Bauman and M. Oguru, 1999. "An Investigation of
Thermal Comfort at High Humidities," ASHRAE Trans. 105(2):94-103.
Forrester, J.R. and W.J. Wepfer, 1984. "Formulation of a Load Prediction Algorithm for
a Large Commercial Building," ASHRAE Trans. 90(2) paper KC-84-09, no.3, Atlanta.
Gabel, S.D., L. Carmichael and G. Shavit, 1998. "Automated Control in Response to

Real-Time pricing of Electricity," ASHRAE Journal p. 26, November.
Gagge, A.P., A.P. Fobelets and L.G. Berglund, 1986. "A Standard Predictive Index of the
Human Response to the Thermal Environment," ASHRAE Trans 92(2):709-731
GenOpt Manual, 2001. http://simulationresearch.lbl.gov/GO/index.html.
Gibson, G.L. and T.T. Kraft, 1993. "Electrical Demand Prediction Using Artificial
Neural Network Technology," ASHRAE Journal p. 60, March.
Goldman, R.F., 1999. "Extrapolating ASHRAE's Comfort Model," Int. J. of HVAC&R

Research 5(3): 189-194.
Gordon, J.M. and K.C. Ng, 2000. Cool Thermodynamics, Cambridge Press, Cambridge,
UK.
Haberl, J.S. and S. Thamilseran. 1996. "The Great Energy Predictor Shootout II:
Measuring Retrofit Savings - Overview and Discussion of Results," ASHRAE Trans.
102(2).
Haberl, J., D. Claridge, D. O'Neal, W. Heffington and W.D. Turner. 1990. "Monitoring
$98 Million in Energy Efficiency Retrofits: The Texas LoanSTAR Program,"
187
Proceedings of the ACEEE Summer Study on Energy Efficiency in Buildings, ACEEE,
Washington, D.C.
Hackner, R.J., J.W. Mitchell and W.A. Beckman, 1984. "HVAC System Dynamics and
Energy Use in Existing Buildings- Part I and H," ASHRAE Trans. 90(2) paper #KC-84-
09,no.2,RP321.
Hamalainen, R.P., J. Mantysaari, J. Ruusunen, and P.O. Pineau, 2000. "Cooperative

consumers in a deregulated electricity market- dynamic consumption strategies and price
coordination," Energy - The International Journal 25:857-875.
Hartman, T.B., 1996. "Design Issues of Variable Chilled -Water Flow through Chillers,"
ASHRAE Trans. 102(2):679.
Haves, P. 2002. "General Guidelines For Emergency Energy Reduction In Commercial

Office Buildings." Lawrence Berkeley National Laboratory, Berkeley, CA.
Henze, G. P., R. H. Dodier, and M. Krarti. 1997. "Development of a Predictive Optimal

Controller for Thermal Energy Storage Systems," Int. J. ofHVAC&R Research 3(3):233-
264.
Ho, K.-L., Y.-Y. Hsu, C.-F. Chen, T.-E. Lee, C.-C. Liang, T.-S. Lai, and K.-K. Chen.
1990. "Short-Term Load Forecasting of Taiwan Power System Using a Knowledge-
Based Expert System," IEEE Trans, on Power Systems, 5(4): 1214-1220.
Ho, K.-L. and Y.-Y. Hsu. 1992. "Short Term Load Forecasting Using a Multilayer
Neural Network with an Adaptive Learning Algorithm," IEEE Transactions on Power
Systems, 7(1): 141-149.
Houck, C.R., J. Joines and M. Kay. 1996. "A Genetic Algorithm for Function
Optimization: A Matlab Implementation." Submitted to ACM Trans, on Mathematical
Software.
Houck, C.R., J.A. Joines and M. Kay. 1998. "Binary and Real-Valued Simulation
Evolution for Matlab GAOT V2." Copyrighted free software that can be redistributed or
modified under the terms of the GNU General Public License as published by the Free
Software Foundation.
Huang, J. and E. Franconi. 1999. "Commercial Heating and Cooling Loads Component
Analysis," Lawrence Berkeley National Laboratory report LBL-37208, Berkeley, CA.
Hydeman, M. 1997, "Water-Cooled Chiller Performance Evaluation Tool Version 2.0,"

PG&E Energy Center, San Francisco, CA.
Iljima, M., K. Takagi, R. Takeuchi, and T. Matsumoto. 1994. "A Piecewise-Linear

Regression on the ASHRAE Time-Series Data," ASHRAE Trans. 100(2): 1088-95.
188
IPMVP, 2001. "International Performance Measurement and Verification Protocol,"
Department of Energy, DOE/GO-102001-1187, Washington D.C., January.
Kahn, E., 1988. Electric Utility Planning and Regulation, American Council for an
Energy Efficient Economy, Washington D.C. and Universitywide Energy Research
Group, University of California, CA.
Katipamula, S. and D.E. Claridge. 1993. "Use of Simplified Systems Model to Measure
Retrofit Energy Savings," ASME Journal of Solar Energy Eng. 115(2):57-68.
Katipamula, S., T.A. Reddy, T.A. and D.E. Claridge. 1998. "Multivariate Regression
Modeling", ASME Journal of Solar Energy Eng. 120:177, August.
Katipamula , S., R.G. Pratt, D.P. Chassin, Z.T. Taylor, K. Gowri and M.R. Bramley,
1999. "Automated Fault Detection and Diagnostics for Outdoor-Air Ventilation Systems
and Economizers: Methodology and Results of Field Testing," ASHRAE Trans. 105(1).
Kawashima, M. 1994. "Artificial Neural Network Backpropagation Model with Three-

Phase Annealing Developed for the Building Energy Predictor Shootout." ASHRAE
Trans. 100(2): 1096-1103.
Kawashima M., C.E. Dorgan and J.W. Mitchell, 1995. "Hourly Thermal Load Prediction
for the Next 24 Hours by ARJJVIA, EWMA, LR, and an artificial neural network,"
ASHRAE Trans. V.101(l), 3849.
Kawashima, M., C.E. Dorgan and J.W. Mitchell. 1996. "Optimizing System Control
with Load Prediction by Neural Networks for an Ice-Storage System," ASHRAE Trans.
102(1):1169-1178.
Kintner-Meyer, M. and J.D. Burns, 2001a. "Utility/Customer Information Services: Part

1: Description of Services and Discussion of Interoperability for service
implementation," ASHRAE Trans. 107(1):306-13.
Kintner-Meyer, M. and J.D. Burns, 2001b. "Utility/Customer Information Services: Part

2: Data Object Modeling and Mapping to BACnet," ASHRAE Trans. 107(l):314-25.
Kissock, J.K., T.A. Reddy, T.A. and D.E. Claridge. 1998. "Ambient-temperature
Regression Analysis for Estimating Retrofit Savings in Commercial Buildings," ASME
Journal of Solar Energy Eng. 120(3): 168-176.
Klein, S.A., W.A. Beckman and J.A. Duffie. 1976. "TRNSYS - A Transient Simulation
Program." ASHRAE Trans. 82(2).
Knebel, D. 1983. Simplified Energy Analysis Method. American Society of Heating,

Refrigerating and Air-Conditioning Engineers, Atlanta.
189
Kreider, J.F. and X.A. Wang. 1991. "Artificial Neural Networks Demonstrated for
Automated Generation of Energy Use Predictors for Commercial Buildings," ASHRAE
Trans. 97(1).
Kreider, J.F. and J. Haberl. 1994. "Predicting Hourly Building Energy Usage: The Great
Predictor Shootout - Overview and Discussion of Results," ASHRAE Trans.
100(2): 1104-18.
LBL, 1980. DOE-2 User's Guide, Version 2.1, Lawrence Berkeley Laboratory and Los
Alamos National Laboratory, Report no. LBL-8689 Rev.2, DOE-2 User Coordination
Office, LBL, Berkeley, CA.
Lee, J.M. and D. Craven, 2001. "Maximizing Building Performance," Buildings

Magazine, July.
Lee, K.Y., Y.T. Cha, and J.H. Park. 1992. "Short-Term Load Forecasting Using an
Artificial Neural Network." IEEE Trans, on Power Systems, 7(1): 124-131. February.
Limaye, D.R. and C. Whitmore, 1984. "Selected Statistical methods for Analysis of Load
Research Data," EPRI Report EA-3467, Research Project 1816-1, May.
Liu, M. and D.E. Claridge. 1995. "Application of Calibrated HVAC System Models to
Identify Component Malfunction and to Optimize the Operation and Control Schedules",
Proceedings of the ASME/JSME/JSES International Solar Energy Conference 1:209,
Maui.
Lorenzetti, D.M. and L.K. Norford. 1993. "Pressure Reset Control of Variable Air
Volume Ventilation Systems," Proceedings of the ASME Solar Engineering Conference,
p. 445, April, Washington, D.C.
Luo, D., L.K. Norford, S.R Shaw and S.B. Leeb, "Monitoring HVAC Equipment
Electrical Loads from a Centralized Location- Methods and Field Test Results," ASHRAE
Trans. 108(2):841-57.
MacKay, D.J.C. 1994. "Bayesian Nonlinear Modeling for the Prediction Competition,"
ASHRAE Trans. 100(2): 1053-62.
MacArthur, J.W., 1986. "Humidity and Predicted-Mean-Vote Based (PMV-based)

Comfort Control," ASHRAE Trans. 92(1B): 5-17.
MacArthur, J.W., A. Mathur and J. Zhao, 1989. "On-Line Recursive Estimation for Load
Profile Prediction," ASHRAE Trans. 95(1), paper CH-89-6-1.
190
Manke, J.M., D.C. Hittle and C.E. Hancock. 1996. "Calibrating Building Energy
Analysis Models using Short-Term Data," Proceedings of the 1996 International ASME
Solar Energy Conference, p.369, San Antonio, TX.
Martin, R.A., C.C. Federspiel and D.A. Auslander, 2002. "Responding to Thermal
Sensation Complaints in Buildings," ASHRAE Trans. 108(1) paper 4536, Atlanta
Mathworks. 2001. Matlab Version 6.1 Release 12.1. Natick, MA.
Mazzucchi, R.P., 1987. "Commercial Building Energy Use Monitoring for Utility Load
Research," ASHRAE Trans. 93(1), paper NY 87-18-2.
Moghram, I. and S. Rahman. 1989. "Analysis and Evaluation of Five Short-Term Load
Forecasting Techniques," IEEE Transactions on Power Systems, 4(4): 1484-1491.
October.
NEMVP, 1996. North American Energy Measurement and Verification Protocol,

Department of Energy Report DOE/EE-0081, March.
Norford, L.K., R.H. Socolow, E.S. Hsieh and G.V. Spadaro. 1989. "Two-to-one
Discrepancy between Measured and Predicted Performance of a Low-Energy Office
Building: Insights from a Reconciliation based on the DOE-2 Model," Energy and
Buildings 21:121.
Norford, L. K., S.L. Englander and B.J. Wiseley. 1998. "Demonstration Knowledge
Base to Aid Building Operators in Responding to Real-Time-Pricing Electricity Rates,"
ASHRAE Trans. 104(1).
Ohlsson, M.B.O., CO. Peterson, H. Pi, T.S. Rognvaldsson, B.P.W. Soderberg. 1994.
"Predicting System Loads with Artificial Neural Networks - Methods and Results from
the "Great Energy Predictor Shootout." ASHRAE Transactions 100(2): 1063-74.
Orange and Rockland, 1987. Residential Peak Activated Rate: Analysis of Results 1982-
1985", Economic and Rate Research, Orange and Rockland Utilities Inc., NJ, January.
Papalexopoulos, A.D. and T.C. Hesterberg. 1990. "A Regression-Based Approach to

Short-Term System Load Forecasting." IEEE Transactions on Power Systems,
5(4): 1535-1544. November.
Park, C , D.R. Clark, and G.E. Kelly. 1985. "An Overview of HVACSIM-I-, a Dynamic
Building/HVAC/Control Systems Simulation Program." Proceedings Is' Annual
Building Energy Simulation Conference, Seattle, WA.
Park, C , D.R. Clark, and G.E. Kelly. 1986. "HVACSIM-i- Building Systems and
Equipment Simulation Program: Building Loads Calculation." NBSIR 86-3331.
National Institute of Standards and Technology, Gaithersburg, MD.
191
Phelan, J., M.J. Brandemuehl, and M. Krarti. 1997a. "In-Situ Performance Testing of
Fans and Pumps for Energy Analysis," ASHRAE Trans. 103(1) paper 4040 (RP-827).
Phelan, J., M.J. Brandemuehl and M. Krarti. 1997b. "In-Situ Performance Testing of
Chillers for Energy Analysis," ASHRAE Trans. 103(1), paper 4040 (RP-827).
Pindyck, R.S. and D.L. Rubinfeld. 1981. Economic Models and Economic Forecasts,
2nd Edition, McGraw-Hill, New York.
Reddy, T.A., J.K. Kissock and D.K. Ruch. 1998. "Uncertainty in Baseline Regression
Modeling and in Determination of Retrofit Savings," ASME Journal of Solar Energy
Eng. 120(3): 185-92.
Reddy, T.A., J. Elleson, J.S. Haberl and D.E. Claridge. 1998. "Literature Review,
Preliminary Methodology Description and Final Site Selection," Progress report for
ASHRAE Research Project 1004, "Determining Long-Term Performance of Cool
Storage Systems from Short-Term Tests," February.
Reddy, T.A. and D.E. Claridge, 2000. "Uncertainty in Measured Energy Savings from
Statistical Baseline Models," Int. J. of HVAC&R Research, 6(2):3-20.
Reddy, T.A., J.S. Elleson and J.S. Haberl, 2002 "Methodology Development for
Determining Long-Term Performance of Cool Storage Systems from Short-Term tests,"
ASHRAE Trans. 108(1) AC-02-19-3 (RP1004).
Reddy, T.A. and K.K. Andersen, 2002. An Evaluation of Classical Steady-State Off-
Line Linear Parameter Estimation Methods Applied to Chiller Performance Data," Int. J.
of HVAC&R Research 8(1): 101-24.
Ried, D., 1987. "Demand-Side Planning Programs: Projects and Products, 1974-86,"
EPRI Special Report EM-5062-SR, February, Palo Alto, CA.
Rumelhard, D.E., J.L. McClelland, and the PDP Research Group. 1986. Parallel
Distributed Processing, Vol.1, Foundations, chapter 8. Cambridge, MA: MIT Press.
Schmitt, D.D., S.A. Klein and D.T. Reindl, 2000. "Automated Generation of Hourly
Design Sequences," ASHRAE Trans. 106(2) paper 4332 (RP-962).
Seem, J.E. and J.E. Braun. 1991. "Adaptive Methods for Real-Time Forecasting of
Building Electrical Demand." ASHRAE Trans. Vol. 97(1):710.
Specht, D.F. 1988. "Probabilistic Neural Networks for Classification, Mapping, or

Associative Memory." Proc. IEEE Int. Conf Neural Networks, Vol.1, pp.525-32, June.
Specht, D.F. 1990. "Probabilistic Neural Networks." Neural Networks Vol 3, pp.109-18.
192
Specht, D.F. 1991. "A general regression neural network." IEEE Trans, on Neural
Networks, Vol .2, pp.568-76, June.
Stevenson, W.J. 1994. "Using Artificial Neural Nets to Predict Building Energy
Parameters," ASHRAE Trans. 100(2): 1081-87.
Thumann, A. (editor), 1999. Customer Choice: Purchasing Energy in a Deregulated

Market, Fairmont Press, Lilburn, GA.
Train, K., J. Hedges, R. Windle, T. Alereza and S. Braithwait, 1985. "Combining

Engineering and Statistical Approaches to Estimate End-Use Load Shapes," EPRI Report
EA-4310, Research project 2145-3, October.
USDOE, 1996. "Measurement and Verification Guidelines for Federal Energy

Management Projects," Department of Energy DOE/GO-10096-248, Washington D.C.
USEPA, 1995. "Conservation and Verification Protocols, Version 2.0," US

Environmental Protection Agency, EPA 430/B-95-012, Washington D.C.
USUI, 1996. "Guidelines for a USUI Project Proposal," US Initiative on Joint

Implementation, US General Accounting Office, Washington D.C.
Usoro, P.B. and I.C. Schick. 1986. "Residential End-Use Load-Shape Estimation, vol.1:
Methodology and Results of Statistical Disaggregation from Whole-House Metered
Loads", Electric Power Research Institute (EPRI) Research Project 2145-1, EM-4525,
Palo Alto, CA.
Vine, E. and J. Sathaye, 1999. "Guidelines for the Monitoring, Evaluation, Reporting,
Verification, and Certification of Energy-Efficiency Projects for Climate Change
Mitigation," Lawrence Berkeley National Laboratory Report LBNL-41543, March.
Williams, L.J., 1979. "New Methods for Short-term Electric Utility Sales Forecasting,"
Symposium Proceedings "How Electric Utilities Forecast," EPRI Report EA-1035-SR,
Special Study Project WS-77-46, March
World Bank. 1994. "Incorporating Social Assessment and Participation into Biodiversity
Conservation Projects," Report by World Bank, Washington D.C.
Wright, J.A. 1996. "HVAC Optimization Studies: Sizing by Genetic Algorithm,"

Building Services Engineering Research and Technology, 17(1):7-14.
Wright, J. and R. Farmani. 2001. "The Simultaneous Optimization of Building Fabric

Construction, HVAC System Size, and the Plant Control Strategy," Proceedings of the
Seventh International IBPSA Conference, 865-872, Rio de Janeiro, Brazil.
193
Wright, J. and H. Loosemore. 2001. 'The Multicriterion Optimization of Building
Thermal Design and Control," Proceedings of the Seventh International IBPSA
Conference, 873-880, Rio de Janeiro, Brazil.
Yang, H.-T. and C.-M. Huang. 1998. "A New Short Term Load Forecasting Approach
Using Self-Organizing Fuzzy ARMAX Models." IEEE Trans, on Power Systems,
13(l):217-225..
Yoshida, H. and T. Terai, 1992. "Modeling of Weather Data by Time Series Analysis for
Air-Conditioning Load Calculations," ASHRAE Trans. 98(1) paper 3576..
Zajac, A. J. 1997 Building Environments, HVAC Systems. Milwaukee, WI: Johnson

Controls, Inc.
194
APPENDIX A OPTIMIZATION TOOL
A. 1 Simulation Model
This appendix describes the prototype optimization tool developed as part of this project.
It includes a description of the simulation model, the set of simulation cases, the structure
of the simulation files, the elements of the objective function, and how the air
conditioning and lighting are controlled.
A.1.1 Model
The simulation model was adapted from a simulation package developed by MTT
graduate student Henry Spindler to explore optimal scheduling of fans and chillers for
night ventilation of buildings. His simulation method can be characterized by:
• Simulation over a single day.

• A 2R2C nodal model of a single room, sized at 100 m2. Temperatures are
calculated for the room air and furniture and for the structural mass of the room.
Solar and internal gains are split between the two temperature nodes, representing
a split between convection and radiation.
• Infiltration from buoyancy-driven airflows.
• Sinusoidal outside temperature, generated from used-supplied maximum and
minimum values, peaking at 3 p.m. and including random noise.
• Solar gains based on a peak amplitude of 600 W/m2 and a sinusoidal amplitude
variation constrained to positive values.
• Internal gains that account for occupants, equipment, and lighting. The building
is unoccupied between 7 p.m. and 7 a.m. and has four occupants, rated at 70
W/person, at 7-8 a.m., 10 occupants 8 a.m. through noon, six occupants for a
lunch hour, 10 occupants until 5 p.m., and five occupants until 7 p.m. Equipment
and lighting loads are lumped to total 40 W/m2. The load fractions are 0.2 from 7
p.m. to 7 a.m., 0.6 at 7-8 a.m., 1.0 at 8 a.m.-5 p.m., and 0.8 at 5-7 p.m.
• An indoor temperature set point that varies over the simulated day.
• Temperature control provided by a proportional-integral or a neural-net controller.
• Cooling provided by open windows alone, open windows and an exhaust fan, and
closed windows and an air-conditioner. Window openings and the operation of
the fan and the air conditioner were optimized by simulating different schedules
and searching for an optimum with a genetic algorithm. The objective function
was a weighted sum of energy consumption and deviation of temperature from the
set point, with the temperature expressed in pseudo-energy units.
The genetic algorithm (GA) was free software (Houck et al. 1996, 1998) and was readily
implemented within the simulation environment, a major benefit.
195
A.1.2 Set of Simulation Cases
The base-case set of simulations focused on reduction of lights and air-conditioning

output. Each service - lights and air-conditioning - was allowed to be restricted to a
value less than required to meet service demands for one of the four afternoon hours.
Allowable values were 0-1.0, in increments of 0.2. This yielded six allowable values for
four hours, or 24 choices for lights and the same number for air conditioning. Because
the two are independent, the total number of possibilities was the square of 24, or 576.
There were three outputs for each simulation: total electricity usage over the four-hour
period, the cumulative deviation of the indoor temperature from the set point, in degree-
hours (positive deviation only), and the indoor-temperature peak.
The outputs were stored in an array, history_array_new. Each row of this array contained
information about a single simulation. A sample is as follows:
2.0000 3.0000 0 4.0000 25.0000 6.0000 32.9407
where the first pair specifies the fractional air-conditioner use (multiplied by 10) and the
hour when the air conditioning is reduced from its full value (hour 3); the second pair
provides the same information for the lights; and the last three figures are the cooling
energy (kWh), degree-hour deviation of temperature relative to the set point, and the
maximum indoor temperature.
Other features of the simulation set included:
• Limitation of the period of interest for load control to the four afternoon hours,
noon-4 pm. The simulation period remained as 24 hours, with specified initial air
and mass temperatures.
• Use of air conditioning only, with closed windows and no exhaust fan. The air-
conditioning system was modeled as a black box of fixed efficiency.
• Summation of all energy used by air conditioning, lights and equipment. This is
important when considering a reduction in lights, to capture the direct benefit of
reduced lighting electricity usage as well as the indirect reduction in heat load and
air conditioning.
A. 1.3 Simulation Structure
The simulation files are of two types: code files and data files, as listed in Tables A.l and
A.2. The code files are available from the Principal Investigators upon request. Figure
A.2 shows the organization of the simulation code files.
196
file purpose
Calc_Deriv Calls Plant
EvaluatePIControl Implements control settings (calculates PI control, prevents
integral wind-up, and limits air conditioning output during the
four afternoon load reduction hours to the values determined
by the function m_initializega) and calculates energy
consumption and cumulative energy consumption
GASearch Sets the number of generations and calls the genetic
algorithm
HourlyEnergy Calculates energy consumption during each of the four load-
control hours
Infiltration_Flow Specifies buoyancy-driven airflow, as a function of window
area, the distance between the lower and upper windows, and
the indoor-outdoor temperature difference
Mechanical_Schedule Specifies whether the chiller and fan are on for three time
periods, comprising hours 0-7, 8-18, and 19-24
my_initializega Creates a random initial population for a genetic algorithm; as
applied to the current research, it creates an exhaustive list of
possible simulations. Each member of a population set is
characterized by a time (one of four hours during the control
period) and a control value, for both lights and air
conditioning.
myode45 Modified solver for ordinary differential equations, altered to
reduce the number of solution steps within a single time step
NetControllerBC Steps through a 24-hour simulation period in one-minute
increments; calls EvaluatePIControl and the ODE solver; and
generates screen plots, at hourly simulation intervals, of
tracking performance, tracking error, controller action, energy
consumption and cumulative energy consumption.
ode45 Solver for ordinary differential equations
Plant Specifies room physical parameters and dynamic
conservation-of-energy equations for two lumped nodes,
representing the room air and lightweight material and the
structural mass
Qint Specifies hourly heat gains from occupants, lights and
equipment
Qsun Specifies solar gain as a sinusoidal function, limited to
positive values
query_history_array Used in my_initializega to establish a population for GA
optimization
results Analyzes energy and temperature output from simulations
Save_data Called at the end of each time step, this function saves the
indoor-air temperature and the HVAC control signal
Schedule_optimization This is the main program, opened in the editor/debugger
window and then run. It uses the function my_initalizega to
197
create a population of test cases, calls SimEval, saves results
in history_array_new, calls GASearch, notes the total number
of simulations, and simulates the best schedule found by the
GA
set_target Sets an indoor-air temperature target, for simulations that
require tracking the target over a diurnal cycle
SimEval Sets initial air and mass temperatures; specifies schedules
(length of simulation, start and stop times, and control actions
for lights and air conditioning during the four-hour control
period, where the control actions are determined by
my_initializega); plots the best and current schedules for air
conditioning and lights; calls the Simulation function; and
weights the three outputs (energy usage, cumulative
temperature deviation, and peak temperature) to form a single
objective-function value suitable for the GA
Simulation Specifies the size of the heating equipment, the chiller and the
fan; sets the time step (one minute for the current research)
and establishes a time array whose elements span a 24-hour
period in 60-second intervals; sets boundary-condition arrays
(outside temperature via the function Tout and solar radiation,
via the function Qsun); calls NetControllerBC; and tabulates
energy consumption, the high-temperature error, and the
maximum indoor temperature
Tout Specifies outside temperature as a sinusoid fit to user-selected
maximum and minimum values, with day-to-day variation
added in the form of a 2 °C sinusoid with a longer period
Table A. 1. Simulation files
File Purpose
Best_Schedule The control schedule that gives the lowest value of the
objective function, as determined by the GA
history_array_new Stores specifications and results for the simulation of each
member of the initial population
rand_vec Vector of random numbers used to represent measurement
noise and to generate a random population for the genetic
algorithm
Table A. 2. Data files
198
Schedule_Optimization
my_initializega GASearch
SimEval
Simu ation
Tout Qsun
ScreenOutput EvaluatePIControl myode45
Mechanical Schedule
1
Calc Deriv
Plant Qsun Qint Tout
I
Infiltration Flow
Figure A.2. Organization of simulation files.
199
A.2 Extensions to the simulation files
The simulation files were incrementally extended in several steps:
• Base case simulation of daily energy consumption of cooling plant;

• Inclusion of lights and equipment in daily energy consumption;
• Calculation of hourly energy use during the four-hour load-control period and
extension of the objective function to include peak demand;
• Addition of thermostat set point as a control element;
• Addition of supply-air temperature for a VAV system as a control element.
Simulation code for each step was retained, making it possible to tailor the set of
simulation files for the problem at hand. Table A.3 shows the files needed for each type
of simulation. An overview of significant features of each type of simulation follows this
table.
200
Base case Lights and equipment Hourly energy use Thermostat Supply-air temperature
control control
Schedule_Optimization Schedule_Optimization Schedule_OptimizationHrEn Schedule_ Schedule.
OptimizationHrEnT OptimizationHrEnTTsa
my_initializega my_initializega my_initializegaHrEn my_initializegaHrEnT my_initializegaHrEnTTsa
SimEval SimEval SimEvalHrEn SimEvalHrEnT SimEvalHrEnTTsa
Simulation Simulation SimulationHrEn SimulationHrEn SimulationHrEn
Set switch to call Set switch to call Set switch to call Set switch to call
NetControllerBCLtEq NetControllerBCLtEq NetControllerBCLtEq NetControllerBCLtEq
NetControllerBC NetControllerBCLtEq NetControllerBCLtEq NetControllerBCLtEq NetControllerBCLtEq
Set switch to call
EvaluatePIControlLtEq VA V
EvaluatePIControl EvaluatePIControlLtEq EvaluatePIControlLtEq EvaluatePIControlLtEq EvaluatePIControlLtEq VAV
Set switch to call Set switch to call Set switch to call
HourlyEnergy HourlyEnergy HourlyEnergy
Calc_Deriv Calc_Deriv Calc_Deriv Calc_Deriv Calc_Deriv
Set switch to call Set switch to call Set switch to call Set switch to call
QintLtEq QintLtEq QintLtEq QintLtEq
Qint QintLtEq QintLtEq QintLtEq QintLtEq
HourlyEnergy HourlyEnergy HourlyEnergy
GASearch GASearch GASearchHrEn GASearchHrEn GASearchHrEn
Results Results ResultsHrEn ResultsHrEn ResultsHrEn
Set switch to analyze a single Set switch to analyze a single Set switch to analyze a single
building or two-building building or two-building building or two-building
combinations combinations combinations
OneBuilding OneBuilding OneBuilding
TwoBuildingCombos TwoBuildingCombos TwoBuildingCombos
Table A3. Simulation files required for each type of simulation.
201
A.2.1 Base Case
As noted above, the base case includes three elements in the objective function,
calculated as follows:
Diurnal energy consumption
Minute-by-minute energy consumption is added by EvaluatePIControl to the

Energy_Consumption vector and to the Cumulative_Energy_Consumption vector. The
function Simulation sums all rows of Cumulative_Energy_Consumption to determine the
daily energy use, Energy_Cons. Table A.4 shows the time scales at which simulation
functions work (daily or minute-by-minute) and the calculation of daily energy
consumption.
.mfile Timing and energy reporting

Schedule Optimization No timing or energy reporting
my_initializega Calls SimEval for a 24-hour simulation
SimEval Calls Simulation for a 24-hour simulation
Simulation Calls NetControllerBC for a 24-hour
simulation on one-minute time steps; sums
all rows of the final column of
Cumulative_Energy_Consumption to
determine the daily energy use,
Energy_Cons.
NetControllerBC Calls EvaluatePIControl, the ODE solver
and Calc_Deriv each one-minute time step.
Plots energy and cumulative energy on a
one-minute basis
EvaluatePIControl Each minute, adds a new element, the one-
minute energy consumption, to the
Energy_Consumption vector and adds a
new element to the
Cumulative_Energy_Consumption vector.
Table A.4. Simulation timing and tabulation of energy consumption.
Cumulative temperature deviation
Minute-by-minute deviations are screened in NetControllerBC and only those where the
actual temperature exceeds the set point are kept and summed. The 24-hour total is then
normalized to obtain units in °C hours.
202
Maximum temperature
The function Simulation identifies the maximum temperature in the array y_array
computed in NetControllerBC.
Base case simulations make use of two control elements, for lights and air conditioning,
determined as follows:
Lighting
Lighting is scheduled in the .m file Qint, to follow a pattern that shows a maximum value
between 8 a.m. and 5 p.m. and lesser values before and after working hours. The variable
light-frac, which varies from 0 to 1.0 in increments of 0.2 and is applied to one hour
between noon and 4 p.m., multiplies the normal value.
Air conditioning
For air conditioning, the variable AC_frac serves to clip rather than multiply the normal
output of the air conditioner. This variable, ranging from 0 to 1.0 in increments of 0.2,
serves as a ceiling for the air-conditioner output calculated for a given hour during the
four-hour load-control period. Table A.5 shows in detail how the output of the air-
conditioner is limited during one of the four load-control hours via the variable AC_frac.
.mfile Air-conditioner control

my_initializega Sets up population
SimEval Defines AC_frac and AC_hour and inserts
into Schedule array
Simulation Calls NetControllerBC and returns
U_vec(:,num_steps) to zero
NetControllerBC Calls EvaluatePIControl
EvaluatePIControl Defines maximum control, u_max, as the
value of AC_frac inserted into Schedule;
defines u as the minimum of u_max and
the control calculated from PI gains;
assigns that control, u, to U_vec for the
time period (minute) under calculation
Calc_Deriv Calls Plant with control u equal to the
element of U_vec for the given time period
Plant Defines AC_setting as the control, u,
passed from CalcJDeriv; defines u_actual
as the product of AC_setting and the
capacity of the cooling plant
Table A.5. Simulation files that limit air conditioning during peak-load hours.
203
A.2.2 Calculation of energy used by lights and equipment
The base case controls lights but does not include lights in the calculation of energy
consumption. It is suitable for investigations of plant energy, which varies with reduction
of air conditioning. Air conditioning energy is reduced via direct control of the
equipment and via a reduction in cooling load due to a reduction in lights. It is
reasonable to optimize total energy usage rather than plant energy, because the former
drives the operating cost of the building. To that end, the second set of simulation
functions includes the energy used by lights and equipment in the objective function, as
documented in Table A.6.
HVAC energy calculation HVAC+lights+equipment

Qint QintLtEq
light_frac, which ranges from 0.2 to 1.0 light_frac reduces only lights; Qlightequip
for a single hour in the four-hour peak and Qtot are separately summed
period, reduces both lighting and
equipment loads
Qint(t) is a scalar load that is used in QintLtEq(t) is a two-element row vector
thermal calculations that includes Qlightequip and Qtot
CaIc_Deriv Calc Deriv
Call Qint Call QintLtEq
NetControllerBC NetControllerBCLtEq
Input array uses Qint as a scalar Input array uses Qtot for thermal
calculations
Call EvaluatePIControl for HVAC energy Call EvaluatePIControlLtEq for energy
consumption consumption of HVAC, lights and
equipment
Simulation Simulation
Call NetControllerBC Call NetControllerBCLtEq
EvaluatePIControl EvaluatePIControlLtEq
Calculates hourly and cumulative (over a Calculates hourly and cumulative (over a
24-hour period) air-conditioning energy 24-hour period) energy consumption for
consumption air conditioning, lights and equipment
Table A.6. Simulation files that record energy used by lights and equipment and use total
energy consumption in objective Junctions.
A.2.3 Hourly energy consumption during load-control hours
Utility charges often include a component that penalizes peak demand. The third set of
simulation files sums energy use on an hourly basis during the peak-load period. The
objective function now has seven elements: energy usage for each of the four load-
control hours, daily energy use, cumulative temperature deviation from set point, and the
peak temperature. Table A.7 summarizes the changes to the simulation files.
204
Simulation file Hourly energy reporting
Schedule_OptimizationHrEn Calls my_initializegaHrEn with
SimEvalHrEn as an argument, retrieves
hourly energy consumption values, and
stores them in history_array_new, calls
GASearchHrEn
GASearchHrEn Calls SimEvalHrEn
my_initializega Calls SimEvalHrEn, stores hourly energy
consumption in pop_bounds
SimEvalHrEn Calls SimulationHrEn
SimulationHrEn Passes Hourly_Energy_Cons to
SimEvalHrEn
NetControllerBCLtEq Calls EvaluatePIControlLtEq
EvaluatePIControlLtEq Calls HourlyEnergy
HourlyEnergy Calculates energy consumption for each of
the four load-control hours, forms the row
vector Hourly_Energy_Cons,
Table A.7. Simulation files required to tabulate energy use during load-control hours.
A.2.4 Including room-air temperature set point as a control variable
The room-air temperature set point was included as a control variable, specifying a single
value of the thermostat set point for each of the four load-control hours.
Schedule_OptimatizationHrEnT specifies bounds on the extent of the variation in room-

air temperature as well as the hours and values for light and air-conditioner control. The
temperature set point is allowed to increase up to 3 °C above the nominal value of 25 °C.
An increase is enforced for all four load-control hours. Schedule_OptimizationHrEnT
calls my_initializegaT, with the file SimEvalHrEnT as an argument.
The function my_initializegaT creates the first three members of the population as before,
but adds values for the thermostat boost. It also creates an exhaustive population with
2304 members. For each of the 576 members previously defined, the thermostat boost is
allowed to range from 0 to 3 °C. This exhaustive search should be used with care due to
the required simulation time: calculating the objective function for a population of 576
required about nine hours on a 450 MHz PHI computer.
The function SimEvalHrEnT creates a schedule for the load-control period that includes
columns for the time, AC control, lighting control, and a fourth column for the thermostat
boost, TstatBoost, which is the same for the four hours. SimEvalHrEnT calls
SimulationHrEn, which in turn calls set_target. The function set_target adds TstatBoost
to the nominal target of 25 °C. Details of the changes to simulation functions are given in
Table A.8.
205
Simulation file Room-air temperature as control
variable
Schedule_OptimizationHrEnT Vector "bounds" includes range of
variation for room-air temperature; calls to
my_initializegaHrEnT and SimEvalHrEnT
my_initializegaHrEnT Establishes the first three members of a
population and includes values for
variation in room-air temperature;
generates exhaustive population
SimEvalHrEnT Establishes a load-control schedule that
includes a four-hour variation in room-air
temperature; calls SimulationHrEn
SimulationHrEn Calls set_target
set_target Adds TstatBoost to the nominal target
temperature
Table A.8. Simulation files required to include room-air temperature as a control

variable.
A.2.5 Modeling a VAV system
Supply-air temperature set point is the last variable to be added to the list of those
controlled to reduce energy costs. This requires two major actions: first, inclusion of
supply-air temperature in the load-control schedule, in a manner similar to that used to
include room-air temperature, and second, a model of a secondary HVAC system.
Schedule_OptimatizationHrEnTTsa specifies bounds on the hours during which the

supply-air set point can be varied and the extent of the variation as well as the hours and
values for light and air-conditioner control and control of the room-air temperature. The
supply-air temperature set point is allowed to increase up to 4 °C above the nominal value
of 13 °C. Note that both the room-air temperature and the supply-air temperature are
allowed to vary in magnitude and in load-control hour when the set points are increased.
Schedule_OptimizationHrEnTTsa calls my_initializegaTTsa, with the file
SimEvalHrEnTTsa as an argument.
The function my_initializegaTTsa creates the first three members of the population as
before, but adds values for the supply-air temperature boost. It also creates an exhaustive
population with 184,320 members. For each of the 576 members of the exhaustive
population for lighting and air-conditioning control there are 320 choices for room-air
and supply-air temperature control: the room-air temperature boost is allowed to range
from 0 to 3 oC over each of four hours (16 combinations) and the supply-air temperature
boost is allowed to range from 0 to 4 °C over each of four hours (20 combinations). This
exhaustive search, if executed, would require a prohibitively long computation period,
about four months on a 450 MHz PHI computer.
206
The function SimEvalHrEnTTsa creates a schedule for the load-control period that
includes a fifth column for the thermostat boost, TsaBoost. Both TstatBoost and
TsaBoost are limited to one of the four load-control hours.
A VAV system was selected for this project as the basis for modeling the secondary
HVAC system. Steady-state equations from the ASHRAE Simplified Energy Analysis
Method (Knebel 1983) were coded into the procedure EvaluatePIControlLtEqVAV. This
procedure accounts for sensible loads only.
EvaluatePIControlLtEqVAV, like EvaluatePIControlLtEq, calculates a control output

based on the deviation of the room-air temperature from the set point and the controller
gains. For the VAV system, the controller regulates airflow to the space and not the
output of the air conditioner. The return-air temperature and volumetric flow and the
outdoor-air temperature and flow then determine the mixed-air temperature and the load
on the cooling coil, which affects chiller power. Fan power is separately calculated and
added to the hourly and daily totals.
Details of the changes to simulation functions are given in Table A.9.
Simulation file Supply-air temperature as control

variable
Schedule_OptimizationHrEnTTsa Vector "bounds" includes range of
variation for Tstat and Tsa; calls to
my_initializegaHrEnTTsa and
SimEvalHrEnTTsa
my_initializegaHrEnTTsa Establishes first three members of a
population with values for variation in
Tstat Tsa; generates exhaustive population
SimEvalHrEnTTsa Establishes a load-control schedule that
includes a one-hour variation in Tstat and
Tsa; defines TsaBoost as a global variable;
and calls SimulationHrEn
SimulationHrEn Calls set_target and NetControllerBCLtEq
set_target Adds TstatBoost to the nominal target
temperature
NetControllerBCLtEq Calls EvaluatePIControlLtEqVAV
EvaluatePIControlLtEqVAV Specifies base value of supply-air
temperature and calculates airflow, fan
power, cooling coil load and chiller power
Table A.9. Simulation files required to include supply-air temperature as a control

variable.
207
A.3 Operation of the simulation environment
The simulation environment is easy to use. For each set of files, described in the
previous section, a simulation is initiated by running Schedule_Optimization or variations
derived from it and denoted with the appended HrEn, HrEnT, or HrEnTTsa. There are
two main issues associated with using the programs: switching between an exhaustive
and a GA search and analyzing and displaying output.
A.3.1 Switching between an exhaustive search and a GA search
The simulation files can be used in two ways. First, they can explore the entire search
space through an exhaustive search. This search systematically simulates all possible
combinations of air-conditioning and lighting control, subject to the restriction that both
can vary over six fractional steps, from 0 to 1.0. Because the control steps are discrete
rather than continuous, the search is more properly denoted as "near-exhaustive." The
second approach generates a relatively small initial population and then uses a GA to
generate additional simulation cases. Table A. 10 summarizes how to switch between
these two modes.
Simulation file Variable or Exhaustive GA search

procedure search
Schedule_Optimization optim_dimension 3 1
Schedule_OptimizationHrEn, 7
HrEnT or HrEnTTsa
num_in_pop 576 for HrEn; User choice;
2304 for HrEnT; investigator
184320 for typically chose a
HrEnTTsa small number (3-
10)
GASearch Delete by
comment sign %
GASearch num_of_gen - User choice
my_initializega Loop counter to ii=l:l ii=l:0
generate
exhaustive
population
SimEval Objective Three-element Single,
function, val value weighted-sum
objective
SimEvalHrEn, HrEnT, and Seven-element function; user-
HrEnTTsa value specified weights
Table A. 10. Variable changes required to switch between an exhaustive search and a GA
search.
208
A.3.2 Simulation output
The GA search identifies which set of control inputs produced the smallest value of the
objective function for a given set of weights. In the first generation, simulations are
conducted for each member of the user-specified population size. The first three
members of the population are based on the control actions and hours specified in
my_initializega. The remaining members have control actions and times selected
randomly. The GA controls the control actions and times for members of successive
generations. For each simulation the schedule of control actions for each of the four
load-control hours and the value of the objective function are displayed in the command
window of the simulation environment. Output also includes simulation times, hourly
plots of temperatures, energy use and control action, plots of the objective functions, and
a final simulation of the best case identified by the GA.
The exhaustive search simply runs through all members of the population set up in
my_initializega. For each, the command window displays the schedule and the elements
(three or seven) of the objective function. Because the number of simulations may be
large, it is not efficient to manually scan the command window to find the best
simulation. Several files were written to display results: results, resultsHrEn, and
OneBuilding and TwoBuildingCombos. The results file works with a three-element
objective function that includes daily and not hourly energy use. It plots cumulative
temperature deviation against daily energy use and produces histograms of daily energy
use, a weighted sum of energy use and cumulative temperature deviation, and a weighted
sum of energy use, cumulative temperature deviation and maximum temperature. The
resultsHrEn file works with a seven-element objective function. It plots cumulative
temperature deviation against daily energy use and calls either OneBuilding or
TwoBuildingCombos. OneBuilding produces the same three histograms as the results
file and also searches for the optimal case when a demand charge is included in the
objective-function weights. TwoBuildingCombos combines output arrays for two
identical buildings, searches for the minimum objective function, and plots a histogram of
objective function values.
A.4. Single-Building Optimization
The case considered in detail in this research included only two control actions over a
four-hour period: reduction of air conditioning for one of the four hours and reduction of
lights for a single hour. Even for this relatively simple situation there are 576
combinations. A richer set of control actions would include a temporary increase in
thermostat set point or adjustments in supply-air or chilled-water temperature. The
number of combinations grows rapidly with the number of possible control actions and
the hours over which they could be implemented, as noted in the previous chapter.
The search for an optimum in a single building was developed in several steps:
1. Use of the code that accounts for the electricity used by lights and equipment;
209
2. Exhaustive search of all possible cases;
3. Genetic-algorithm search;
4. Use of the code that accounts for hourly energy use during the peak period and a
peak-electricity charge;
5. Exhaustive search of all possible cases;
6. Genetic-algorithm search.
A.4.1. Single-case simulations, with electricity used by air

conditioning, lights and equipment
In this first step, the simulation code was run for the five cases shown in Table A.l 1. The
first three were a standard set generated as the first members of all populations to be
examined with the GA. The fourth run was a repeat of the second run, but with much
lower control gains. The purpose of these runs was to establish some confidence in the
simulation and to gain expectation for reasonable outputs. The third case - turning off
the air conditioning and the lights during the last of the four load-control hours - was
repeated in run five, when the schedule was such that the air conditioning remained off
for the remainder of the day. For all runs, the code included lights and equipment in the
daily energy use.
Table A. 11 shows that daily energy use was 80.8 kWh when there was no reduction in
service. There was very little cumulative increase in temperature above the set point of
25 °C: the maximum temperature was 25.3 °C. For a sense of scale, the peak power for
lights and equipment was 4 kW (20 W/m2 each, and 100 m2 floor area), which would
yield 96 kWh if they were fully used throughout the day and there were no air-
conditioning load. Energy use declined 2.47-2.49 kWh when the air conditioning and
lights were fully turned off during the second or fourth load-control hours. Turning the
lights off completely for a single hour reduced electricity usage by 2 kWh. Turning off
the air conditioning provided a further reduction during the load-control hour but
generated a pick-up load the following hour. There was very little difference in energy
consumption, and temperature deviations as well, when air conditioning and lights were
turned off during hour 13 or during hour 15.
The maximum temperatures for runs 2 and 3 rose 8.3-8.5 °C above set point. The
cumulative temperatures increased above set point about 8 °C hours, indicating that the
temperature was quickly returned to set point after air conditioning was restored. (A
temperature that rises to 8 °C above set point at the end of one hour and returns to set
point at the end of a second hour gives approximately 8 °C hours of cumulative
temperature deviation.) Lower control gains, in run 4, yielded miniscule increases in
energy consumption and temperature deviation from set point. When the air conditioning
was controlled to turn off during hour 15 and was allowed to stay off for the remainder of
the day (run 5), energy use plummeted but the temperature was poorly controlled.
210
Run AC AC Lights Lights Energy Cumulative Maximum
hour control hour control kWh temperature Temperature
deviation °C
°C hour
1 Nordn Nordn 80.8158 0.3456 25.3299
2 13 0 13 0 78.3280 7.7964 33.8198
3 15 0 15 0 78.3470 7.5741 33.5806
4 (run 2 but with low 13 0 13 0 78.3378 8.1304 33.8373
gains)
5 (run 3 but with no 15-24 0 15 0 67.8265 73.1771 37.0378
A/C after hour 15)
Table A.ll. Daily electricity usage, cumulative temperature deviation and maximum
temperature for four variations of control for lights and air conditioning.
Figures A.3-A.7 show plots of temperatures, control action and energy usage for the five
runs. Each figure has five plots. The top plot shows the sinusoidal outside temperature,
varying from 12 to 28.5 °C, the indoor set point, a constant 25 °C, the measured indoor
temperature (indoor temperature with measurement noise added), and the temperature of
the building mass. The single simulation day is not repeated; as a result, the air and mass
temperatures at the end of the day may differ from those at the beginning, when both
match the indoor-air the set point. In particular, the mass temperature drifts slowly
upward over the day, in part due to radiant loads, while the air temperature, is either right
at set point (Figure 4.1) or shows a rapid change during and following the load-control
hour.
The second plot is the tracking error, in °C. This plot has two lines. One, with
oscillations, is the difference between the target and the measured indoor-air
temperatures. The second line is the tracking error smoothed with an exponential moving
average; the simulation code for air-conditioner controller can make use of either and the
latter was used throughout this project.
The last three plots show the air-conditioner controller action, electrical power in kW,
and cumulative energy consumption in kWh. Air-conditioning control and output is zero
until internal gains pick up at 7 a.m. and the tracking error becomes negative. The
mechanical schedule permits the air conditioner to run as needed throughout the 24-hour
day. At night, the outside temperature is low and the small internal gains provide heat to
balance the conductive loss. The result for this particular combination of heat losses and
gains is the same as would be achieved if the air conditioner were forced off at night.
The control action also shows significant oscillation. Lower gains produce a more
damped response, as shown in Figure A.7.
Power drops by about 4 kW, as shown in Figure A.4, when the lights and air conditioning
are turned off. Two kW is due to lights and the remainder to the air conditioner, which
has a peak electrical power of 3.5 kW. When the air conditioner is turned on again and
the controller saturates, the air conditioner is running at peak and the total power is 7.5
kW, with 2 kW for lights, 2 kW for equipment, and 3.5 kW for the air conditioner. The
211
peak load for the day is therefore higher when air conditioning is reduced at some point
and restored, than when it is operated normally.
Tracking Performance
30
20
Cumulaîge Energy Consumption kWh 20 25

100 ! !"
50
m 1 ' i
10 15 20 25
Figure A3. Temperature tracking, controller action, and energy use for a case where
there is no reduction in lights and air conditioning.
212
40
20
10 Tracking Error 15 20 25
10
-10
10 Controller Action 15 20 25
0.5
1 glectrical Power kVfe 20 25

10
5 -
Cumulate Energy Consumption kWh 20 25

100
50
10 15 20 25
Figure A.4. Temperature tracking, controller action, and energy use for a case where
there is full reduction in lights and air conditioning during hour 13.
213
-10
1 o Controller Action \ 5 20 25
0.5-
1 Electrical Power kVfe 20 25

10

100
50-
10 15 20 25
there is full reduction in lights and air conditioning during hour 15.
214
-10
10 Controller Action 15 20 25
0.5
1§lectrical Power kVYs 20 25

10

100
50
10 15 20 25
there is full reduction in lights and air conditioning during hour 13 and the controller
gains are reduced to produce smaller controller and temperature oscillations.
215
-20
1 glectrical Power kVYs
Cumulate Energy Consumption kWh

100
Figure A. 7. Temperature tracking, controller action, and energy use for a case where
there is full reduction in lights and air conditioning during hour 15 and the air
conditioner is kept off for the remainder of the day.
A.4.2. Exhaustive search of all possible cases
The simulation code was modified to perform an exhaustive search, identifying and
simulating all 576 possible cases. Each case was simulated over a day. Each simulation
required about one minute and the set required about nine hours, on a relatively slow (333
MHz PUJ) computer. Two full sets of simulations were performed. In the first set, the
maximum air-conditioning power established for the last of the load-control hours was
used for the remainder of the day. In the second set, the maximum power was returned to
full output after the load-control period.
For the first set, energy use and temperature deviation were plotted in two dimensions, as
shown in Figure A.8. The maximum energy consumption was about 20% higher than the
minimum. Most of the simulations show relatively high energy use and low temperature
deviation. These are associated with air-conditioning control that did not curtail service
after the load-control period. There are three V-shaped, smaller clusters of points in the
216
lower right (just above the set of six near-vertical strands), in center of the figure, and in
the upper left. These are associated with reduced air conditioning in the late afternoon
and evening. The cluster in the lower right is due to a maximum air-conditioning output
of 0.4 for hours 15-24, the cluster in the middle is due to a limit of 0.2, and the cluster in
the upper left, with the highest thermal penalty, is due to full shutdown of the air
conditioner.
Cumulative temperature deviation vs. energy consumption ove r a 24-hour day

1 1 i 1 1 1 1
66 68 70 72 74 76 78 80 82
electricity consumption (kWh) •
Figure A. 8. Cumulative temperature deviation and daily electricity consumption for 576
simulations. Lowest energy cases correspond to highest temperature deviation, when the
air conditioning was fully turned off for hour 15 and the remainder of the day.
As shown in Figure A.9, the same type of plot looks very different when the control was
changed such that air conditioning resumed operation after being shut down during hour
15. The range of energy usage is much smaller in Figure A.9, which in essence expands
the lower right region of Figure A.8.
It is worthwhile to examine Figure A.9 closely, looking at both horizontal and vertical
patterns. Notice the regular horizontal spacing between the six slightly tilted vertical
bands. These bands correspond to reduction of lights from full on to entirely off, in
increments of 0.2. A previously noted, full lighting power is 2 kW. A reduction of 0.2
217
corresponds to 0.4 kW. The spacing is slightly larger, just in excess of 0.5 kW. The
difference is the reduction in air conditioning load associated with a decrease in lights.
The air conditioner COP is 3 and a reduction in thermal load of 0.4 kW will reduce air
conditioning by about 0.13 kW. Moving vertically along one of the six strands of data
points corresponds to a very slight (a fraction of one percent) increase in energy
consumption as the air conditioning is limited to a fraction of its maximum value, with
the fraction varying from 0 to 1 in increments of 0.2. Turning the air conditioning
completely off for a single hour causes the largest increase in temperature deviation and
very slightly boosts energy usage, due to the pick-up load the following hour and the
gains on the temperature controller.
Cumulative temperature deviation vs. energy consumption ove r a 24-hour day
79 79.5 80 81
electricity consumption (kWh)
Figure A.9. Cumulative temperature deviation and daily electricity consumption for 576
simulations. Lowest energy cases correspond to highest temperature deviation. If turned
off during hour 15, air conditioning resumes operation for the remainder of the day.
Daily electricity usage, in kWh, cumulative temperature deviation, in °C hours, and

maximum temperature were then weighted to form three different scalar objective
functions. The relative weights for the three objective functions were (1,0, 0), meaning
that only energy was considered; (1,1,0), weighting energy and cumulative temperature
deviation equally; and (1, 1, 0.1), weighting the maximum temperature at 0.1. The
218
objective function values were then binned in histograms, shown in Figures A.10-A.12,
for cases when air conditioning remained off if completely curtailed during hour 15, and
in Figures A.13-A.15, for cases when air conditioning curtailment was limited to the four
load-control hours and did not extend from hour 15 through the remainder of the day.
Figure A. 10 shows that maximum energy consumption was about 20% higher than the
minimum and that most of the cases were at the high end. This is simply a presentation,
in one dimension, of the information in Figure A.8: most of the runs did not substantially
reduce air-conditioning energy usage. Figure A.l 1 bins the weighted sum of energy and
cumulative temperature deviation and shows that most of the cases were at the lower,
more optimal values. These cases correspond to relatively high energy use and low
temperature deviation. The very worst cases, with an objective-function value greater
than 140, are those in the upper left of Figure A.8. When the temperature deviation was
weighted as heavily as energy use, a control strategy that turned off the air conditioning
at hour 15 and left it off was a poor choice. Figure A. 12 includes a weight for maximum
temperature as well as cumulative temperature deviation and is similar in appearance to
Figure A.l 1.
219
Frequency of occurence of daily energy usage cases
300 i i i i i i
250 ^ ^ ^ | -
200 -
150 ^^^B -
100 ^ ^ ^ | ^^^B -
50 ^^^H -
n ^^^^^^^^^H ^^^^^^^^^^L^^BHB^^^^^^^^^H
66 68 70 72 74 76 78 80 82
Figure A. 10. Histogram of occurrences of daily energy consumption among all cases
considered, when the maximum output of the air conditioning at hour 15 was set as the
maximum for the remainder of the day.
220
Frequency of occurence of a weighted sum of energy and temp erature deviation
4001 , 1 1 1 1 1 1
70 80 90 100 110 120 130 140 150
Figure A.11. Histogram of weighted sum of electricity usage and cumulative temperature
deviation, when the maximum output of the air conditioning at hour 15 was set as the
maximum for the remainder of the day.
221
Frequency of occurence of a weighted sum of energy, tempera ture deviation and maximum temperature
80 90 100 110 120 130 140 150 160
Figure A. 12. Histogram of weighted sum of electricity usage, cumulative temperature

deviation, and maximum temperature, when the maximum output of the air conditioning
at hour 15 was set as the maximum for the remainder of the day.
Figure A. 13 shows the same narrow range of energy consumption as in Figure A.9 and a
relatively uniform spread across the range. It is not fruitful to try to read too much into
the pattern, which is a function of the size of the bins and how they map onto the energy
consumption values shown in Figure A.9. For example, the relatively low number of
cases with energy consumption of 80.5 kWh is due to the discrete gap in energy values
caused by a change of 0.2 in fractional lighting power.
Figure A. 14 is more interesting, because temperature deviations are penalized. The range
in the objective function, a weighted sum of energy and cumulative temperature
deviation, was larger, about 15%. When temperature deviation is important and can be
quantitatively valued, it is important to pick a control strategy that accounts for this
deviation. The distribution shows that there were a large number of cases with
essentially the same optimal performance, suggesting that it should not be difficult to find
one without having to search exhaustively. Figure A. 15 shows a pattern similar to Figure
A.M. With maximum temperature also weighted, the distribution peaked slightly more
sharply at optimal values.
222
Frequency of occurence of daily energy usage cases
100
Figure A. 13. Histogram of occurrences of daily energy consumption among all cases
considered.
223
Frequency of occurence of a weighted sum of energy and temp erature deviation
i i i i i i
20
00
^H
- ^ ^ .
80
60
40
^^M
^ ^ ^ ^ ^ ^ ^ ^ ^ _ ^ -
20
78
^^^ÎH 80 82 84 88
Figure A. 14. Histogram of occurrences of weighted sum of daily energy consumption and
90 92
cumulative temperature deviation.
224
Frequency of occurence of a weighted sum of energy, tempera ture deviation and maximum temperature
1401 r
120-
100
Figure A. 15. Histogram of occurrences of weighted sum of daily energy consumption,

cumulative temperature deviation, and maximum temperature.
A.4.3. Genetic-algorithm search
The purpose of the GA search is to find a near-optimal solution in less time than is
required by an exhaustive search. This section documents the results of a number of
runs, over which were varied the population size, the number of generations of the
search, and the weights on cumulative temperature deviation and absolute temperature.
Runs were performed in two sets, as for the exhaustive search. In the first set, the
maximum output of the air conditioning at hour 15 was set as the maximum for the
remainder of the day. In the second set, the air conditioning resumed normal operation
after being reduced during the last load-control hour.
In all runs for both sets, the first three members of the population of the first generation
were not random but were deliberately seeded. They had the same control strategies as
were used in the first three cases in Table A.ll: no reduction in lights and air
conditioning, full reduction in the second load-control hour (hour 13) and full reduction
in the last load-control hour (hour 15).
225
Table A. 12 shows the results from the first set of runs. The runs are clustered in groups
of four, with each group assigned the same weights. Population size and number of
generations varies within each group. Run 8 was repeated to determine whether the
stochastic nature of the search would lead to different results, which it did not. Table
A. 12 shows the number of simulations for each run, which is often less than the product
of population size and number of generations. This is because the code looked for cases
that had been simulated in a previous generation and did not duplicate the simulation, as a
means of speeding up the search.
Across different groups there was substantial difference in objective functions, because
the weights change. But note that the best objective-function value for each run showed
very little variation within a group. Six simulations did not consistently find the best
case, as is shown in Table A. 12 for run 5. The best simulation should keep the lights
fully off for one hour, because a reduction in lighting power is not penalized. Sets of 30,
however, were essentially the equal of sets of 100. It has already been noted that there
were a large number of cases that offer similar performance for given weights.
For the first cluster of runs, runs 1-4, where there was no thermal-comfort penalty, the
GA consistently selected the control strategy that turned off the air conditioning during
the fourth load-control hour and left it off. There was a slight decrease in power, 0.53
kWh, when the lights were turned off in hour 12 rather than hour 15. In the former case,
both lighting and air conditioning loads dropped in hour 12 and air conditioning later
dropped in hours 15-24. In the latter case, the air conditioning and lights were switched
off at the same time. In runs 5-12, with thermal penalties as determined by the selected
weights, there was little or no reduction of air conditioning. As was seen in the
exhaustive search, very little energy was saved when the air conditioning was reduced.
The results in Table A. 13, for the second set of runs, are similar. Values of the objective
function changed relative to those in Table A. 12, because the air-conditioner control was
different. There are several interesting points. First, for runs l-4a, the same objective
function was reported for three different cases when the lights were fully turned off in
hour 12: run 2 (no air-conditioner reduction), run 4 (air conditioner limited to 0.8) and
run 4a (air conditioner limited to 0.73). This is because the air-conditioner control is an
upper limit: if the load during hour 12 is less than the limit, there is no reduction.
Second, run 5, which used the minimum set of six simulations, found a best case with
noticeably more energy than the other runs with the same weights, runs 6-8. However,
the percentage increase was still small, about 2.3%. Third, four runs, including three
with the minimum of six simulations, identified an optimal case that did not have the
lights fully turned off, indicating that the search space was not sufficiently large.
Note that the results in Tables A. 12 and A. 13 are essentially the same when there was a
thermal-comfort penalty, in the form of a weight on cumulative temperature deviation
alone or both the cumulative deviation and the maximum temperature. Restricting air-
3
While the exhaustive search limited service reductions to increments of 0.2, the GA search space included
service reductions in increments of 0.1.
226
conditioning use for hours 15-24 was attractive (and gave lower energy usages in Table
A.12 than in Table A.13) only when energy consumption alone constituted the objective
function.
run pop gen num Lights Obj.
of
Temp A/C A/C Lights
sim deviation hour control hour control function
weight
1 3 2 6 0 0 15-24 0 15 0 67.8265
2 3 10 26 0 0 15-24 0 12 0 67.2930
3 10 2 16 0 0 15-24 0 15 0 67.8265
4 10 10 62 0 0 15-24 0 12 0 67.2930
5 3 2 5 0 Nordn 12 0.1 78.8822
6 3 10 27 0 Nordn 13 0 78.6230
7 10 2 18 0 Nordn 13 0 78.6230
8 10 10 65 0 13 0.9 13 0 78.6230
8a 10 10 63 0 15 0.8 13 0 78.6230
9 3 2 5 0.1 Nordn Nordn 83.6943
10 3 10 28 0.1 15 0.9 13 0 81.1560
11 10 2 20 0.1 Nordn 15 0 81.1677
12 10 10 58 0.1 15 0.7 13 0 81.1560
Table A.12. GA search for optimal control schedule, as a function of population size,
number of generations and weighting function. Air conditioning turned off in hour 15
remained off for the rest of the day. "No rdn" means that there was no reduction in
service.
run pop gen num Temp Max A/C A/C Lights Lights Obj.
sim deviation temp hour control hour control function
weight weight
1 3 2 6 0 0 13 0 13 0 78.3280
2 3 10 29 0 0 Nordn 12 0 78.2080
3 10 2 17 0 0 13 0.9 13 0 78.2136
4 10 10 66 0 0 12 0.8 12 0 78.2080
4a 10 10 70 0 0 12 0.7 12 0 78.2080
5 3 2 6 0 Nordn 15 0.7 80.3911
6 3 10 26 0 Nordn 12 0 78.6292
7 10 2 19 0 Nordn 15 0 78.6348
8 10 10 57 0 15 0.9 13 0 78.6230
9 3 2 6 0.1 Nordn Nordn 83.6943
9a 3 2 6 0.1 12 0.9 12 0.9 83.4399
10 3 10 29 0.1 12 0.8 13 0 81.1560
10a 3 10 26 0.1 Nordn 15 0 81.1677
10b 3 10 27 0.1 13 0.7 13 0 81.1560
11 10 2 17 0.1 14 0.7 13 0.3 81.9160
12 10 10 63 0.1 Nordn 13 0 81.1560
12a 10 10 69 0.1 Nordn 13 0 81.1560
Table A. 13. GA search for optimal control schedule, as a function of population size,
number of generations and weighting function.
227
From Tables A.12 and A.13, consider the minimum value of the objective functions for
runs 1-4, when energy only was considered in the objective function used with the GA,
runs 5-8, where temperature deviation and maximum temperatures were assigned weights
of 1 and 0, respectively, and Runs 9-12, where the weights were 1 and 0.1. These
objective-function values can be compared with those found in an exhaustive search of
all possible cases, as shown in Table A. 14. This table shows that when the objective
function was energy alone, the GA found the best case, as identified by the exhaustive
search. For the cases where temperature was included in the objective function, the GA
came close enough for practical purposes but did not exactly match the best case found
via the exhaustive search. Consider the last row of Table A. 14. Of the 576 cases
examined by the exhaustive search, eight had an objective-function value of 81.1560 and
two had the minimum values of 81.1539. In contrast with the best GA control strategies
in runs 12 and 12a of Table A. 13, one of the two optimal runs identified via the
exhaustive search capped the air conditioning at 0.6 for hour 14 and completely turned
off the lights during the same hour.
A/C control, Temperature Maximum GA Search Exhaustive Search

if off hour 15 deviation weight temperature
weight
Off 15-24 0 0 67.2930 67.2930
Off 15-24 1 0 78.6230 78.6209
Off 15-24 1 0.1 81.1560 81.1539
Off 15 only 0 0 78.2080 78.2080
Off 15 only 1 0 78.6230 78.6209
Off 15 only 1 0.1 81.1560 81.1539
Table A. 14. Comparison of the minimum value of the objective function determined by
the GA with the value found in an exhaustive search.
Figures A.16-A.19 show the objective functions for all of the simulations included in
each generation of four GA searches, runs 9-12 in Table A. 13. Note that there were
many cases that had similarly low values of the objective function (smallest negative
number is desired). Figure A. 19 shows fewer simulations as the number of generations
increases: the GA homed in on a near-optimal solution and found cases that have already
been simulated.
228
Plot of all generations evaluated
-83 i i i i i
-
+ +
-84
-85 - -
-86 - -
-87 - -
-88 -
+
-89 -
+
+
-90 - -
-91 -
-92
+
Q1 i i i i
0 0.5 1 1.5 2 2.5
Figure A. 16. All simulations for a GA search with a population size of three, two
generations, and weights of 1.0 and 0.1 for cumulative temperature deviation and
maximum temperature.
229
1 1 1 4- 4-
4- 4- 4- 4- 4-
+ + + +
+ +
- + + -
+ + + +
- -
- -
+
+
1 1 1 i i i i i i i
10 11
Figure A. 17. All simulations for a GA search with a population size of three, 10
230
i i
+ +
+ *
* +
+ +
- +
+ +
$
-
+
i i i
0 0.5 1 1.5 2 2.5
Figure A. 18. All simulations for a GA search with a population size of 10, two
231
81
4- 4- 4- 4 - 4 - 4 - 4 - 4- 4- 4-
+ +
+ + +
82 + + + + + + -
+
+ + + +
+ + +
+ +
83
+ ~
+ +
+ + * + +
84
+ + -
+ +
+ +
85 + -
86 - -
+ +
87 -
88 + -
+ +
89 +
+
an i i i i i i i i
10 11
Figure A.19. All simulations for a GA search with a population size of 10, 10
Progress toward the best case among those simulated is shown in Figures A.20 through
A.22. These three runs were all for the same population size, three, and number of
generations, 10. Variation from run to run was due to the stochastic nature of a GA
search. Note that there was essentially no improvement after four generations in Figure
A.20 and after two generations in Figures A.21 and A.22.
232
Trace of the Best and the Avg value achieved
Figure A.20. Progression ofGA search, showing the best and average energy
consumption for each generation. There were three cases in each generation and 10
generations.
233
-87
consumption for each generation. The population size, number of generations and
weights are identical to those in Figure A. 19.
234
-84
-88
consumption for each generation. The population size, number of generations and
weights are identical to those in Figure A. 19.
AAA. Use of simulation environment that includes a compilation of

hourly energy consumption during load-control hours
To include the cost associated with a demand charge in the objective function, it is
necessary to separately sum energy use over time intervals appropriate for calculating
demand. In practice, demand can be set from 15-minute rolling-average electricity
consumption. A demand charge can include blocks, such that the charge per kW depends
on the amount of power. Charge per kW can increase (ascending blocks) or drop
(descending blocks). Whether or not the peak load for a given day contributes to the
monthly or annual demand charge depends on the magnitude of the load relative to that
on other days.
This study used a fixed, 60-minute demand window. Accordingly, energy use was
summed over each of the four load-control hours. The user can set a weight for the
demand; in practice, this could be zero if the loads predicted for a given day fall below
235
those already experienced in the demand period and are likely not to set the largest
demand for a year-long sliding window if charges are based on a ratcheted demand.
A.4.5. Exhaustive search with peak demand included
The output array for an exhaustive search included hourly energy consumption for the
four load-control hours as well as electricity use for the entire day, cumulative
temperature deviation, and maximum temperature - seven elements in all. As before, the
exhaustive search generated 576 cases. One set was run, with air-conditioning under
normal control regardless of reductions during the last load-control hour. Results from
the exhaustive search will be compared with the GA results in the next section.
A.4.6. Genetic-algorithm search with peak demand included
The GA was used as before, to find a near-optimal solution with fewer simulations than
required by the exhaustive search. Results are shown in Table A.15. The check run
included no demand charge in the weighting and gave the same objective-function value
as GA searches performed without hourly energy calculations. The 24 runs were
clustered in six groups of four, each with the same weights. Within each group of four,
the population size and number of generations were varied, as before.
There was very little difference in objective-function values for the four runs in each
group. These runs were set up to require six to 100 simulations, although the actual
number was reduced when the GA found combinations of weights that had been
simulated earlier in the search. Notably, the large runs of ten generations with a
population of ten cases did not require 100 simulations and never needed more than 74.
Six simulations were essentially as good as at least ten times that many. The largest error
due to taking only six simulations was in runs 9-12, when the objective-function for a
population size of three and two generations was about 3% higher than the objective-
function value for a population of ten and ten generations.
A truly optimal solution includes a control combination that turns the lights off fully for
one hour, because there is no penalty associated with reducing lights and it saves energy.
Six of the 24 runs did not identify such combinations. The minimal simulation sets, with
a population size of three and two generations, were most vulnerable to not finding such
combinations, as expected. This accounted for the largest variation in objective-function
values within each group: runs 9-12, as noted above, and runs 13-16.
Air conditioning was turned off in several of the runs that weighted only energy and
demand. When thermal penalties were included, air conditioning was either not reduced
at all or limited to relatively large fractions of full power.
236
lit
run pop gen num Dmd Max A/C A/C Light Light Obj.
lit
sim weight temp hr cntl hr cntl function
weight
chk 3 10 29 0 0 0 12 0.7 12 0 78.2080
1 3 2 6 1 0 0 15 0 15 0 84.7307
2 3 10 28 1 0 0 15 0.5 12 0 84.6027
3 10 2 16 1 0 0 15 0 15 0 84.7307
4 10 10 74 1 0 0 13 0.9 12 0 84.5892
5 3 2 6 10 0 0 15 0 15 0 142.1846
6 3 10 27 10 0 0 15 0.1 13 0 141.8331
7 10 2 16 10 0 0 15 0.9 14 0.1 142.0169
8 10 10 62 10 0 0 14 0 15 0 141.8122
9 3 2 6 1 1 0 No No 87.5554
rdn rdn
10 3 10 25 1 0 12 0.6 12 0 85.0090
11 10 2 19 1 0 No 15 0 85.0185
rdn
12 10 10 63 1 0 No 14 0 84.9857
rdn
13 3 2 6 10 0 No No 145.1020
rdn rdn
14 3 10 28 10 0 No 14 0 142.1396
rdn
15 10 2 16 10 0 13 0.9 12 0 142.4407
16 10 10 63 10 0 12 0.9 14 0 142.1396
17 3 2 6 1 0.1 12 0.6 12 0.4 88.5582
18 3 10 26 1 0.1 12 0.9 14 0 87.5187
19 10 2 19 1 0.1 14 0.6 14 0.1 87.7647
20 10 10 73 1 0.1 14 0.7 14 0 87.5187
21 3 2 6 10 0.1 No 13 0.4 146.0298
rdn
22 3 10 28 10 0.1 14 0.8 14 0 144.6726
23 10 2 20 10 0.1 14 0.9 15 0 145.0054
24 10 10 69 10 0.1 12 0.8 13 0 144.9050
Table A. 15. GA search for optimal control schedule, as a function of population size,
number of generations and weighting function. The weighting function now includes a
weight on electricity demand.
Table A. 16 compares the best objective function found by GA search with the minimum
and maximum values found via the exhaustive search, for different weights. A demand
weight of LOO relative to an energy weight of 1 corresponds to a demand charge of $10
per kW and an energy charge of $0.10. Limiting the demand weight to 10 therefore
underplays the cost of demand. As will be discussed next, this did not affect the
identification of the best strategy.
Table A. 17 provides control strategies for the minimum and maximum objective-function
values. A% expected, failure to reduce light levels resulted in the largest values of the
objective function, because there was no penalty for reduced illuminance and fully
reducing the lights one hour gave the best performance. Also as expected, the minimum
237
and maximum objective-function values tracked the weights: high demand charges
produced significantly higher objective-function values than did lower weights, for a
given set of temperature weights, and temperature weights modestly increased objective-
function values for given demand and energy weights. What was not expected was that
the same control strategy produced the best performance for all six sets of weights, and a
second strategy consistently produced the worst performance. This strongly suggests that
pre-screening possible strategies may simplify the search procedure and avoid elaborate
day-ahead optimization calculations.
Runs Demand Energy Cumulative Maximum GA Exhaustive Exhaustive

weight consumption temperature temperature minimum search search
weight weight weight objective minimum maximum
function objective objective
function function
1-4 1 0 0 84.5892 84.5653 88.4935
5-8 10 0 0 141.8122 141.6756 155.9935
9-12 1 1 0 84.9857 84.9768 98.6615
13-16 10 1 0 142.1396 142.0871 166.1615
17-20 1 1 0.1 87.5187 87.5098 102.2813
21-24 10 1 0.1 144.6726 144.6201 169.7823
Table A. 16. Comparison of the best objective function found by GA search with the
minimum and maximum valuesfound via the exhaustive search, for different weights.
II!
Dmd AC Lgt Min obj Lght Max obj

If!
Temp AC Lgt AC AC Lgt

wght dev hr cntl hr cntl fnc hr cntl hr cntl fnc
wght
1 0 0 14 0.6 14 0 84.5653 14 0 No 88.4935
rdn
10 0 0 14 0.6 14 0 141.6756 14 0 No 155.9935
rdn
1 1 0 14 0.6 14 0 84.9768 14 0 No 98.6615
rdn
10 1 0 14 0.6 14 0 142.0871 14 0 No 166.1615
rdn
1 1 0.1 14 0.6 14 0 87.5098 14 0 No 102.2813
rdn
10 1 0.1 14 0.6 14 0 144.6201 14 0 No 169.7823
rdn
Table A. 17. Control strategies associated with the minimum and maximum objective
function-values found by exhaustive search.
238
A.5. Two-Building Optimization
A.5.1 Possible Combinations of Control Strategies for Two Buildings
The simulation approach for this project was modified in several ways. The most
important change stemmed from a consideration of how best to search for the optimal
control strategy over two or more buildings.
Consider a single building with 1,000 possible control strategies. For two independent
buildings there would be 1,000,000 cases. A GA would search for the optimum in a way
that would be an improvement over an exhaustive search that would require a million
simulations. However, the GA would not easily recognize that the search space consists
of two identical but independent buildings, and that there are only 1,000 different
building simulations to consider. The question is whether the GA would find the
optimum by using fewer than 500 two-building runs, each of which would require two
simulations. If not, a more efficient approach would be to simulate all 1,000 and then
establish an optimal by forming all possible two-building combinations and looking for
the minimal value of the objective function. The argument in favor of the latter approach
becomes more compelling as the size of the problem grows, and this was the approach
selected for the work at hand.
Even with 576 and not 1,000 control cases for a single building, there are 331,776 two-
building combinations. In lieu of examining all of them, two much smaller populations
were investigated, consisting of nine and 100 possible two-building combinations.
Consider the first three members of any population used for a GA search, as shown in
Table A.18. Applying these schedules to two buildings yields nine combinations,
identified in Table A. 19. If the two buildings are identical, there are only six distinct
combinations. In practice, however, any two buildings will differ. The simulation
environment is set up to consider this situation.
Hour Case 1 Case 2 Case 3

AC Light AC Light AC Light
control control control control control control
12 1 1 1 1 1 1
13 1 1 0 0 1 1
14 1 1 1 1 1 1
15 1 1 1 1 0 0
Table A.18. Control schedules for seeded members of population.
239
Combination Building 1 case Building 2 case
1 1 1
2 1 2
3 1 3
4 2 1
5 2 2
6 2 3
7 3 1
8 3 2
9 3 3
7aWe A. 19. Combinations of control actions for two buildings, each of which can be
controlled according to one of three possible schedules.
A.5.2 Simulation Results
Which combination is considered most desirable will depend on the weights applied to
peak demand, energy usage, cumulative temperature deviation and maximum
temperature. Table A.20 shows the minimum value of the objective function for different
sets of weights. As for a single building, a large weight for the maximum electricity
demand produced a large value of the objective function. Further, non-zero weights for
the cumulative temperature deviation and the maximum temperature also boosted the
value of the objective function. When only energy costs were considered (run 1), the best
combination was number 5, for which air conditioning and lights were turned off in both
buildings during hour 13. When demand and energy costs were the only components of
the objective function (runs 2 and 3), the best combination was number 9, in which air
conditioning and lights were turned off during hour 15. This result may appear
surprising, because the presence of a demand charge might favor staggering the hours of
reduced air conditioning. However, a reduction in air conditioning in hour 15 leaves the
increase in load beyond the normal level, in hour 16, outside the period during which
demand charge is computed. When an increase in temperature beyond normal levels was
factored in, via weights on the cumulative temperature deviation or the maximum •
temperature, the best combination was always number 1, for which there is no reduction
in service. In short, the optimum solution was to do nothing to reduce load.
240
Run Demand Electricity Cumulative Maximum Best Minimum
weight consumption temperature temperature Combination Objective
weight deviation weight function
weight
1 0 1 0 0 5 156.6560
2 1 1 0 0 9 169.4614
3 10 1 0 0 9 284.3692
4 0 1 0 1 162.3226
5 1 1 0 1 178.1108
6 10 1 0 1 290.2040
7 0 1 0.1 1 167.3886
8 1 1 0.1 1 180.1767
9 10 1 0.1 1 295.2700
Table A.20. Minimum objective function for a two-building combination, for each of the
nine sets of weights for energy usage and temperature.
Table A.21 shows the variation among the nine possible combinations for each of the
runs in Table A.20. This is interesting, because it indicates the range of costs associated
with each set of weights.
For run 1, the worst option was combination 1, doing nothing, as expected when energy
alone is penalized. Combination 5, the best option, turning off lights and air conditioning
in both buildings in the second of the four load-control hours, yielded an objective-
function value that was smaller by only a miniscule amount than objective-function
values for combinations 6, 8 and 9. These three combinations were the other ways in
which air conditioning and lights were completely turned off for a single hour in both
buildings (combinations 6 and 8 for different hours and combination 9 for the same
hour). The range of values from best to worst was only 3.2%. That it was not larger,
given that one of the options was doing nothing, is due to the increase in air conditioning
load after an hour of no use, to return indoor temperatures to the set point.
Runs 2 and 3 introduced a small and large demand charge, respectively. For run 2, doing
nothing was the worst option. For the larger demand charge, run 3, the worst
combination was number 5, for which air conditioning and lights were turned off during
the same hour. This makes sense, because simultaneously turning off air conditioning
will lead to high loads in both buildings during the next hour. For the higher of two
demand weights, the range from best to worst was about 8%, sufficiently large to make it
worthwhile to identify the best control strategy.
Runs 4-9 introduced penalties on thermal comfort. The results were consistent, in that
doing nothing was best when a reduction in air conditioning increased indoor
temperatures and the associated thermal-comfort penalty. The worst strategy for each run
was again combination 5, turning off air conditioning and lights during hour 13. The
range from best to worst varied with the weights, from 5.1% (run 5) to 10.2% (run 9, with
a high demand charge).
241
Run Combination
1 2 3 4 5 6 7 8 9
1 161.6315 159.1437 159.1627 159.1437 156.6560 156.6749 159.1627 156.6749 156.6939
2 174.4196 173.0378 171.9302 173.0378 171.6560 170.4960 171.9302 170.4960 169.4614
3 289.5129 298.0844 286.8380 298.0844 306.6560 294.8857 286.8380 294.8857 284.3692
4 162.3226 167.2857 167.0824 167.2857 172.2488 172.0455 167.0824 172.0455 171.8421
5 178.1108 181.1798 179.8499 181.1798 187.2488 185.8655 179.8499 185.8655 184.6096
6 290.2040 306.2264 294.7576 306.2264 322.2488 310.2562 294.7576 310.2562 299.5174
7 167.3886 173.2007 172.9734 173.2007 179.0128 178.7855 172.9734 178.7855 178.5582
8 180.1767 187.0948 185.7409 187.0948 194.0128 192.6066 185.7409 192.6066 191.3257
9 295.2700 312.1414 300.6487 312.1414 329.0128 316.9962 300.6487 316.9962 300.2335
Table A.21. Variation in objective-function value for each member of the population and
each set of weights for energy and temperature. The minimum and maximum values are
highlighted, the former in bold and the later in bold italics.
Next, consider a larger population, of 10 rather than three. The first three were the same
as previously used while the remaining seven were selected at random. The goal here
was to determine whether there were combinations that yielded lower values of the
objective function. Ten cases chosen for a particular run are shown in Table A.22. These
ten cases were combined to form 100 possible two-building cases. The first ten matched
the first member of the population set for Building 1 with each of the ten for Building 2,
the second ten matched the second member of the population set for Building 1 with each
of the ten for Building 2, and so on. The lowest values of the objective functions for the
same series of weights as were used with a population of three are displayed in Table
A.23.
Note that only four of the 100 possible combinations gave minimum objective-function
values for the nine different runs: combinations 34, 35 (selected three times), 20, and 45
(selected four times). With energy and demand weights alone, the best combinations
were 34, 35, and, with high demand, 20. Combinations 34 and 35 featured shutting off
the lights and air conditioning in hour 15 (case 3) for one building. For the other
building, case 4 also turned off the lights in hour 15 and limited the air conditioning to
50% of maximum output in hour 14. Such a limit yielded a modest reduction in energy
for that hour but an increase in the following hour. Case 5 reduced the lights to 0.1 of
required output in hour 13 and limited the air conditioning to 0.9 of maximum in hour 12.
This small reduction in maximum allowable powers saved essentially nothing. Required
air-conditioning power was less than the limit.
Combination 20 had air conditioning and lights shut down completely in hour 13 (case 2)
in one building and a modest reduction in air conditioning in hour 15 and substantial
reduction in lights (to 0.2) in hour 14 in the other building. Staggered reductions in lights
reduced peak demand charges.
When thermal-comfort penalties were included, only combinations 35 and 45 were

selected as best. Again, staggered lighting reduction was at the core of the preferred
242
strategies. Thermal penalties made it less likely that substantial reduction in air
conditioning would be considered optimal.
It was fruitless to try to analyze why one particular control combination is the best, given
that the search space featured a number of combinations with very similar values of the
objective function. For example, Figure A.23, a histogram of objective-function values
for case 9, shows that a third of the 100 combinations are in the lowest bin.
Case AC hour AC Lights Lights

control hour control
1 Nordn Nordn
2 13 0 13 0
3 15 0 15 0
4 14 0.5 15 0
5 12 0.9 13 0.1
6 13 0.8 13 0.2
7 15 0.5 15 0.2
8 15 0 14 0.6
9 13 0.9 12 0.2
10 15 0.4 14 0.2
Table A.22. Ten cases generated for investigation of 100 possible two-building
combinations.
Run Demand Electricity Cumulative Maximum Best Minimum

weight consumption temperature temperature Combination Objective
weight deviation weight function
weight
1 0 0 0 34 156.4987
2 1 0 0 35 169.1874
3 10 0 0 20 281.3387
4 0 0 45 157.7511
5 1 0 45 170.5161
6 10 0 35 284.4612
7 0 0.1 45 162.8171
8 1 0.1 45 175.5821
9 10 0.1 35 289.7956
Table A.23. Minimum objective-function values for a single-building population size of

10.
Table A.24 compares the minima found with populations of three (nine combinations)
and 10 (100 combinations). The differences are small, ranging from 0.10% to 4.45%.
This suggests that, again, it is not necessary to have an extremely large search for the
optimum, given the small number of control actions (reduction of air conditioning and
lighting during only one of four hours). The control strategies considered best for the two
searches may have been very different, but their effects were essentially the same. For
this reason, no search was conducted for the minimum among all possible two-building
combinations.
243
Frequency of occurence of objective function
351 1————i 1 1 1 1 r
30-
25 -
20-
15-
10-
5 -
285 290 295 300 305 310 315 320 325 330
Figure A.23. Histogram of occurrences of objective-function value for run 9.
Run Minimum Minimum Difference,

objective objective (9 combo - 1 0 0
function, 9 function, 100 combo)/100
combinations combinations combo *100%
1 156.6560 156.4987 0.10
2 169.4614 169.1874 0.16
3 284.3692 281.3387 1.08
4 162.3226 157.7511 2.90
5 178.1108 170.5161 4.45
6 290.2040 284.4612 2.02
7 167.3886 162.8171 2.81
8 180.1767 175.5821 2.62
9 295.2700 289.7956 1.89
Table A.24. Comparison of objective-function minima for single-building populations of

three and 10, which give two building populations of nine and 100.
244

Rp-1146-Building Operation and Dynamic Within An Aggregated Load PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Rp-1146-Building Operation and Dynamic Within An Aggregated Load PDF

Uploaded by

Copyright:

Available Formats

Shaping Tomorrow’s

Built Environment Today

Building Operation and Dynamics

American Society of Heating, Refrigeration and Air-Conditioning Engineers

T. Agami Reddy Civil & Architectural Engineering, Drexel University

Leslie K. Norford Tabors Caramanis & Associates

August 30, 2002

Aggregation of Load-Control Customers: This set of four tools is proposed to aid an

Jason Lukes, graduate student at Drexel University, Philadelphia

To accomplish these goals, the following discrete tasks were proposed:

Phase 1: Identification of Load Aggregation Opportunities

Phase 2: Load Aggregation Opportunity Assessment

2.1 Electric Utility Aggregation

2.1.1 Aggregate Billing

2.1.3 Building Optimization - Individual Buildings

2.1.4 Load Aggregation - Multiple Buildings

2.2 Historical Perspective

2.2.1 Participation in Deregulation

2.2.2 Restructuring & Retail Access

So far, retail access has been implemented in California, Massachusetts, and

2.2.3 Current Picture of Electric Utility Market

Power Generating Station Independent Generator

LOCAL DISTRIBUTION UTILITY

Figure 2.1. Regulated electric utility market structure.

Power Supplier A , Power Supplier B

Energy Service Provider (ESP)

REGIONAL POWER GRID - TRANSMISSION - INDEPENDENT

Figure 2.2. Electric utility market structure following deregulation.

2.2.4 Rate Components and Metering

The price for a kWh of electricity is comprised of an energy (commodity) charge, a

One of the prerequisites for obtaining service under a demand-measured

Figure 2.3. Demand Metered Commercial Buildings, Aggregated Size Categories.

Demand Metered Commercial Buildings as Percent of

Size, Thousands of Square Feet

According to CBECS Table 10, "Electricity Consumption and Expenditures Intensities,

Number Electricity Used Median Peak Median

Table 2.1. Electricity consumption and demand by building type.

2.2.6 Energy Service Providers and Their Customers

Examples of energy service agreements struck in 1999 include energy management

Allegheny Subsidiary of Volume pricing Large commercial and industrial users.

Table 2.2. Summary of aggregated loads by provider.

3.1 General Criteria for Selection

3.2 Selected Candidate Sites

As required by ASHRAE 1146-RP, six candidate sites were identified as indicated in

Table 3.1. Candidate sites.

3.3 Preliminary Site Interviews

3.3.1 University of Delaware

Electric Utility Information

Contributions to the Project

Electric Utility Information

3.3.3 Fort Dodge Animal Research Facility

Electric Utility Information

It should be noted that under the automatic-supervision project, a significant amount of

Presently, a proposal is being entertained regarding the construction of a 2.6MW

Contributions to the Project

Organization Location Contact Phone

Table 4.1 Case study sites.

4.3 Presentation Format

(1) Metered Group Details

(5) Chart of Metered Group Use

2*' b»i ISJkCtt

Figure 4.1. Metered group summary key.

4.3.1 Metered Group Details