You are on page 1of 9

390 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO.

1, JANUARY 2019

Smoothing Net Load Demand Variations Using


Residential Demand Management
Fadi Elghitani and Ehab El-Saadany , Fellow, IEEE

Abstract—One of the major obstacles facing the large- wi Weighting parameter to prioritize queue i.
scale integration of renewable energy sources to existing Qi (t) Queued energy in queue i at time slot t.
power networks is the large magnitude of fluctuations and Zi (t) Virtual queued energy in virtual queue i at time slot t.
uncertainties introduced to aggregated demand profiles.
Such variations make the supply–demand matching a more ai (t) Arrived energy to queue i at time slot t.
challenging task, and increase the operational cost of the Piu Uncorrelated RES power for consumer i.
power system. In this paper, we provide an effective, yet P0 Correlated RES power.
simple, methodology for smoothing power variations us- r(t) Aggregated net noncontrollable demand at time slot t.
ing the demand response of a large number of residential
θ Temperature controlled by a TCL appliance.
appliances. We first present a demand aggregation model
based on queueing theory that can accommodate both de- θe Ambient (external) temperature.
ferrable and thermostatically controlled loads. Second, con- xi (t) Scheduled (satisfied) energy from queue i at time slot t.
trollable demands are scheduled using an online algorithm
to smooth the net noncontrollable demand profile. The per-
formance of our methodology is evaluated using realistic I. INTRODUCTION
data. LTHOUGH utilizing renewable energy sources (RESs) is
Index Terms—Aggregation model, demand management,
Lyapunov optimization, renewable energy sources (RES),
A not a new idea, there are several challenges for integrating
them in existing power networks in terms of grid stability and
residential appliances. power system operations [1]–[3]. The latter’s challenges are
attributed to output power fluctuations from wind turbines or
NOMENCLATURE photovoltaic cells. These fluctuations are characterized by being
Q Set of queues. stochastic, making the generation nondispatchable, which leads
i Virtual arrived energy to virtual queue i. to two major operational concerns. First, large fluctuations may
θs Temperature threshold for the TCL to send a demand result in a supply–demand imbalance that can contribute to a
request. high ACE [4], [5]. The second concern is to provide adequate
θm ax Maximum required temperature level of a TCL appli- reserves to accommodate the uncertainties in the net demand,
ance. i.e., the remaining demand after reducing the portion satisfied
θm in Minimum required temperature level of a TCL appli- by the RESs. High uncertainties imply more reserves to be
ance. purchased [6], [7], which can lead to a significant increase in
B Uncontrollable tradeoff parameter between general de- the cost of operations.
lay and smoothing performances. Many research studies were made to smooth power fluctua-
Cth Thermal capacitance. tions resulting from the RES integration. One approach, which
N Number of demand classes (queues). avoids the deployment of expensive storage systems [8], is to
PTCL Power rating of a TCL appliance. modify electric energy demand in a way that helps in smoothing
Rth Thermal resistance. power fluctuations. Contributing consumers will be compen-
V Controllable tradeoff parameter between general delay sated using suitable financial incentives. In the literature, this
and smoothing performances. approach is referred to as demand response (DR), which covers
a wide range of applications, other than RES integration. The
Manuscript received January 29, 2018; revised May 29, 2018; ac- majority of the existing solutions are based on resource manage-
cepted June 8, 2018. Date of publication July 2, 2018; date of current ment (optimization) frameworks, in which demand modification
version January 3, 2019. Paper no. TII-18-0242. (Corresponding author:
Fadi Elghitani.)
is the resource to be dispatched. Reducing power fluctuations
F. Elghitani is with the Network Planning Department, National can appear explicitly in the objective function [9], [10], or im-
Telecommunication Institute, Cairo 11768, Egypt (e-mail: fadi.ghitani@ plicitly by assuming a convex cost function of the aggregated
nti.sci.eg).
E. El-Saadany is with the Department of Electrical and Computer
power profile [11], [12]. Existing studies allocate different types
Engineering, University of Waterloo, Waterloo ON N2L 3G1, Canada, on of demands at the beginning of the scheduling time horizon (of-
leave from the Petroleum Institute, Khalifa University, Abu Dhabi 127788, fline scheduling) based on the forecast of the RES output power.
UAE (e-mail: ehab@uwaterloo.ca).
Color versions of one or more of the figures in this paper are available
However, the forecast might not be accurate, especially for long
online at http://ieeexplore.ieee.org. time horizons. Also, offline scheduling is not suitable for man-
Digital Object Identifier 10.1109/TII.2018.2852482 aging residential appliances, because it is hard to predict when
1551-3203 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
ELGHITANI AND EL-SAADANY: SMOOTHING NET LOAD DEMAND VARIATIONS USING RESIDENTIAL DEMAND MANAGEMENT 391

power), we assume that their cost is a quadratic function of the


mismatch. The AS can be directly purchased by the GENCO or
by the DISCO if the latter acts as a balancing responsible party.
The GENCO/DISCO is responsible for compensating the vari-
ations in net demand, which is the total uncontrollable demand
minus the RES generation. To achieve the aforementioned goal,
the GENCO/DISCO is allowed to directly control some of the
appliances of the residential consumers, who will be compen-
sated using suitable financial incentives. For simplicity (and in
consistency with existing studies in the area), we model our
system as a discrete time system, where information is gathered
Fig. 1. System model under consideration. The residential appliances
are classified into three categories: deferrable, TCL, and noncontrollable. and decisions are made in prespecified time slots. The duration
Appliances are controlled by the EMC. The communication is established of each time slot is 1 min.
with the GENCO/DISCO via the smart meter of each consumer. We follow an approach for residential demand modeling sim-
ilar to that in [13]. The general behavior of consumers can
a consumer will exactly need to turn an appliance ON. Finally, be modeled as a Markov chain, whose states represent differ-
demand allocation decisions are usually the solution of a mixed- ent consumer’s activity level. Each activity level can influence
integer linear programming problem, which becomes computa- the probability that a consumer turns ON an appliance. Each
tionally difficult as the number of appliances to be scheduled consumer has three types of appliances: deferrable appliances,
increases. TCLs, and noncontrollable appliances. Both deferrable appli-
In this paper, we propose a methodology for smoothing power ances and TCLs are controllable, i.e., can contribute to DR.
fluctuations resulting from RES integration, using the DR of a An example of deferrable appliances is the washing machine,
large number of residential appliances. Our contributions are whose energy consumption can be deferred freely, but before a
threefold: First, we develop an aggregation model for the resi- deadline dictated by the consumer’s settings. Thus, a deferrable
dential DR. The model exploits different DR potentials of dif- appliance is controlled only by defining how many time slots it
ferent consumers and it is suitable for both deferrable and TCL can be deferred before consuming energy. An example of TCL
appliances. Second, based on the aforementioned model, we appliances is the refrigerator that can be turned ON or OFF freely
develop a DR scheduling algorithm that aims to reduce the as long as the refrigerator’s temperature is within a certain tem-
fluctuations resulting from a large-scale integration of RESs. perature band (the deadband). The last type of appliances are
The algorithm is computationally efficient and avoids solving those that cannot be controlled (e.g., TV), and hence, they have
a large-scale optimization problem. Third, we evaluate the pro- no contribution to DR.
posed algorithm based on two criteria: long-term average perfor- In addition to appliances’ energy demand, each consumer has
mance and short-term sample-path performance. Different from a local RES generator, which satisfies a part of the consumer’s
the existing studies, our proposed methodology is based on an demand. Since RES units are installed within a limited geo-
online scheduling method for residential appliances, which is graphical area, the power profiles generated from all units are
suitable to be applied to any general RES signal. Also, our highly correlated. We model the power generated from the RES
methodology depends on an aggregation model for residential unit of consumer i as a weighted sum of two power profiles
appliances, which can accommodate a large number of appli-
PiRES = αP0 + (1 − α)Piu . (1)
ances.
The remainder of this paper is organized as follows. We de- The first profile represents the nominal power profile P0 re-
scribe the system model in Section II. Our proposed method- sulting from the “spatially average” RES power in the consid-
ology and problem formulation are presented in Section III. ered geographical area. This profile is the same for all RES
Numerical results are given in Section IV to demonstrate the units; hence, it represents the correlated part of the RES power
performance of our proposed methodology. Finally, Section V production. The second profile Piu represents the deviation from
concludes this research. that nominal power profile as a result of any possible imperfec-
tions. The profile depends on the conditions of the corresponding
II. SYSTEM MODEL RES unit and is independent of other units. The overall effect of
the aggregated RES power is the contamination of the smooth
Consider a standalone microgrid as shown in Fig. 1. The mi- aggregated demand profile with fluctuations.
crogrid is composed of a residential community that is served
by a generation company (GENCO). The generation capabil-
ity has a ramping time Ψ > 0, which is the time needed for III. PROBLEM FORMULATION
the generators to reach from minimum to maximum production The demand controller has to schedule a large number of ap-
capability. Any supply–demand mismatch is covered by pur- pliances in real time such that the cost of the purchased AS is
chasing expensive ancillary services (AS). To account for both minimized. Consumer’s convenience should be guaranteed to
regulation-up (AS that are available to increase output power) a high level to ensure the continuity of consumers’ DR partic-
and regulation-down (AS that are available to decrease output ipation. Consumer’s convenience is achieved when deferrable
392 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

where θ and θe are the temperature to be controlled and the


temperature of the TCL’s surroundings, respectively, and Rth ,
Cth , and PTCL are three fixed parameters of the TCL appliance,
representing the thermal resistance, the thermal capacitance,
and the power rating, respectively. The function s(t) is a binary
decision variable that takes either 0 or 1. To achieve consumer’s
convenience, the controlled temperature should be kept within
the deadband

Fig. 2. Demand aggregation model. Each demand request is assigned θm in  θ  θm ax (3)


to the queue of the highest delay class that is lower than the request’s
deadline. where θm in and θm ax represent the minimum and maximum
temperatures, respectively, beyond which the operation of the
loads are scheduled to consume their required energy before the TCL becomes inconvenient for the consumer. Let the appliance
deadline and the temperatures of the TCL appliances are kept send a request for energy demand whenever the temperature
within their deadbands. falls below a certain threshold θs ∈ [θm in , θm ax ], which can be,
As discussed in Section I, it is difficult to use individual de- for example, the temperature set point. The appliance requires
mand models when the number of appliances to be scheduled is an energy consumption E and a deadline D for satisfying the
large. Hence, we propose a demand aggregation methodology demand. Parameters E and D are selected such that the TCL’s
for residential appliances based on the queueing theory to facil- temperature stays within the deadband between two successive
itate the scheduling of controllable residential appliances. First, requests, regardless of the demand controller’s decision.
whenever a controllable appliance needs to consume energy, it After sending the demand request, the appliance is not turned
ON until it receives permission from the EMC. The maximum
sends a demand request to the energy management controller
(EMC), and then waits until it receives an approval from the time that the appliance can wait without causing a discomfort
EMC to be turned ON. The demand request simply consists of for the consumer is the period that θ drops from θs to θm in .
the amount of energy required and the maximum number of Therefore, the longest possible deadline Dm ax can be obtained
time slots the appliance can wait, i.e., a deadline for satisfying from (2) as follows:
 
the request. Similar to [14], we introduce a layer that acts as θs − θe
an interface between different consumer appliances and the de- Dm ax = Rth Cth ln . (4)
θm in − θe
mand controller, referred to as the DR management layer, which
assigns the requests into different queues according to their ur- Since the TCL appliance does not have a prior information
gencies, i.e., the deadlines for satisfying the requests, where when it will receive the permission for power consumption, the
each queue represents a distinct class of demand, as shown duration of the power consumption K should be specified under
in Fig. 2. The set of all possible classes Q = {1, 2, . . . , N } a worst case condition. The TCL power consumption duration
is defined in advance and is the same for all residential should be sufficiently long for the controlled temperature to rise
consumers. above θs . The worst case happens when the appliance receives
the permission exactly at the end of the deadline, when θ is at its
minimum value. Therefore, the minimum power consumption
A. TCLs as Deferrable Loads
duration Km in is determined as the time required for θ to rise
It is relatively easy to generate demand requests from de- from the aforementioned minimum value to θs . The duration
ferrable loads. For example, the energy needed for a washing Km in can be derived from (2) as
machine to complete a washing cycle is known in advance (or  
can easily be deduced). The deadline for satisfying the energy PTCL Rth − (θs − θe )e−D /R th C th
Km in = Rth Cth ln . (5)
request can be adjusted by the consumer. On the other hand, PTCL Rth − (θs − θe )
demand requests from TCLs cannot be generated in the same The minimum required energy for satisfying the TCL’s de-
straightforward manner. For TCLs, consumer’s comfort is de- mand is simply given by
fined by controlling temperatures to be within certain dead-
bands, not by adjustable delay deadlines as in deferrable loads. Em in = PTCL Km in . (6)
Therefore, we need to deduce a method for representing demand
requests of TCL appliances similar to that of deferrable appli- B. Demand Scheduling
ances. Our proposed method is described in the following. First,
We use Lyapunov optimization [16]–[18] to schedule demand
we assume that each TCL follows a simple equivalent thermal
requests. Each queue i ∈ Q is updated according to
parameter model, where the controlled temperature evolves as
follows [15]: Qi (t + 1) = [Qi (t) − xi (t)]+ + ai (t) (7)
θ(t) = θe + s(t)PTCL Rth where Qi (t), xi (t), and ai (t) are the deferred, the scheduled, and
the newly arrived demands of class i at time slot t, respectively.
− (θe + s(t)PTCL Rth − θi )e−t/R th C th (2) The function [x]+ is defined as max{x, 0}.
ELGHITANI AND EL-SAADANY: SMOOTHING NET LOAD DEMAND VARIATIONS USING RESIDENTIAL DEMAND MANAGEMENT 393

In order to take consumers’ comfort into consideration, one


Algorithm 1: Controllable Demand Scheduling Algorithm.
can set different limits (constraints) on queue length, i.e.,
Qi (t)  Qm ax
. However, this approach just provides a soft Input: V ; wi ∀i ∈ Q
i
deadline guarantee, i.e., there can be a nonzero probability that 1: for t ∈ 0, 1, 2, . . . do
a demand request waits beyond the deadline. To provide a hard 2: Measure/read the following variables:
deadline guarantee, we follow the same approach as in [18], r(t), Qi (t), Zi (t) ∀i ∈ Q
where we introduce a set of virtual queues, each is associated 3: Calculate the total optimal scheduling energy xs (t)
with one of the real queues. The virtual queues are updated as by the following equation:
N N
follows:  [Qi (t) + Zi (t)]
i= 1 wi
+
xs (t)  xi (t) = − r(t).
Zi (t + 1) = [Zi (t) − xi (t) + i 1{Q i (t)> 0} ] (8) i= 1
2N V
(10)
where i  E{ai (t)} is a positive constant, and 1{.} is an indi-
4: First individual energy allocation is given by
cator function that equals one when the condition between the
xi (t) = min{xs (t)/N, Qi (t)} ∀i ∈ Q.
braces is satisfied and zero, otherwise. By setting different limits 5: Distribute the remaining energy
on these virtual queues such that Zi (t)  Zim ax , the maximum 
xr (t) = xs (t) − N i=1 xi (t) among queues
waiting time for a request in queue i is given by [18]
arbitrarily wherever possible.
Dim ax = (Qm
i
ax
+ Zim ax )/i . (9) 6: Update Qi (t + 1) according to (7) and Zi (t + 1)
according to (8) ∀i ∈ Q.
On the other hand, the variables {xi (t)} should be controlled
7: end
dynamically to reduce the fluctuations imposed by the noncon-
trollable net demand profile r(t), which represents the resulting
demand after subtracting the portion served by the local RES where Q̄i is the time average of the mean length of queue i.
generation. Our measure for these fluctuations is the time aver- Problem P1 represents an extreme case when consumers are
age of the second moment of the total power consumption willing to wait as long as it takes to achieve the best possi-

2 ⎫ ble performance c∗ . Practically, residential consumers will set
1  ⎨ ⎬
T −1 N
c̄  lim E r(t) + xi (t) . deadlines for scheduling their demand requests, e.g., as in (4).
T →∞ T ⎩ ⎭ The long-term performance of Algorithm 1 is described in
t=0 i=1
the following theorem.
Hence, our goal is to minimize c̄. We will first propose an Theorem 1: Assuming all queues are initially empty,
algorithm for scheduling the controllable demand {xi (t)}. Af- Qi (0) = 0 and Zi (0) = 0 ∀i ∈ Q, and assuming system inputs
terward, the performance of the proposed algorithm will be ω(t) = {r(t); ai (t) ∀i ∈ Q} are i.i.d. and bounded, then ap-
addressed by Theorems 1 and 2. Our proposed algorithm is plying the Algorithm 1 yields the following results.
described in Algorithm 1. The algorithm needs a set of pos- 1) All queues, both real and virtual, are upper bounded by
itive constants {V ; wi ∀i ∈ Q} as an input, which we will Qm ax
and Zim ax , defined by
i
later describe their effect on the performance of Algorithm 1.
2N V
The algorithm is online, where scheduling decisions are made Qm
i
ax
 {N am
i
ax
+ rm ax } + am
i
ax
based on the current realization of the input stochastic processes wi
{r(t); ai (t) ∀i ∈ Q} and on the system’s history demonstrated and
in the variables {Qi (t), Zi (t) ∀i ∈ Q}. The algorithm does 2N V
not need any form of prediction for the stochastic processes Zim ax  {N i + rm ax } + i .
wi
{r(t); ai (t) ∀i ∈ Q}.
Next, the performance of Algorithm 1 is described by Theo- 2) The finite time horizon average cost satisfies

2 ⎫
rems 1 and 2, which will be defined shortly. The first theorem is T
 −1 ⎨ N ⎬
1
related to the long-term average behavior of the cost function, E r(t) + xi (t)  c∗ + B/V
whereas the second theorem is related to the short-term sample T ⎩ ⎭
i=1 i=1
path behavior of the cost function.
where B is a constant given by
We begin by setting a benchmark for performance compari-   
son, c∗ . For the long-term behavior, we define c∗ as the infimum N
(xm ax 2
) + (am ax 2
) max 2i , (xm
i
ax 2
)
i i
of the solution of the following stochastic optimization problem: B wi + .
2 2
i=1
P1 : min c̄ (12)
x i ,∀i∈Q
Proof: See Appendix A. 
s.t. (7)
Theorem 1 describes the long-term average performance. We
T −1
1  will now describe the performance of Algorithm 1 from a sample
Q̄i  lim E {Qi (t)} < ∞ ∀i ∈ Q path perspective. We should first find another measure for perfor-
T →∞ T
t=0 mance comparison other than c∗ . We follow a sample path com-
(11) parison, in which we assume that inputs {r(t); ai (t) ∀i ∈ Q}
394 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

are given in advance for a finite time horizon T . We use this TABLE I
NONCONTROLLABLE APPLIANCES’ PARAMETERS (EXTRACTED FROM
information to define the following deterministic optimization MEASUREMENTS IN [20])
problem:
k T +T −1

N
2
1  
P2 : min ck  r(t) + xi (t)
x i ,∀i∈Q T
t=k T i=1

s.t. (7)
k T
+T −1
[xi (t) − ai (t)]  0 ∀i ∈ Q
t=k T
k T
+T −1 TABLE II
TCL APPLIANCES’ PARAMETERS
[xi (t) − i ]  0 ∀i ∈ Q. (13)
t=k T

We here divided the infinite time horizon into a series of


frames, each of length T , where the scheduling cost of frame k
is given by ck . Hence, our new comparison benchmark will be
the optimal solution of problem P2. Assume we run Algorithm
1 for a number of K frames, then the sample path performance
is given by the following theorem.
Theorem 2: Assuming all queues are initially empty,
Qi (0) = 0 and Zi (0) = 0 ∀i ∈ Q, then applying Algorithm 1 demand data of one of these consumers to represent the demand
for a number of K frames yields the following sample-path behavior of all the consumers in our system. The appliances
result: used in our study are discussed in the following.
K T −1

N
2 K −1
1   1  ∗ 1) Noncontrollable Appliances: We consider three noncon-
r(t) + xi (t)  ck + BT /V. trollable appliances having significant energy consumption,
KT T
t=0 i=1 k =0 which are TV, kettle, and lighting devices. From the measure-
Proof: See Appendix B. ment data, we extract the parameters given in Table I. The pa-
 rameters include the appliances’ power rating, the probability
that an appliance is turned ON at any given time slot, and the
IV. NUMERICAL RESULTS approximating distribution of their power consumption duration
(with its mean given in brackets in units of time slots).
A. Raw Data 2) Deferrable Appliances: We consider two appliances
In our simulation, we assume a single type of RES, which whose demands can be deferred: the washing machine and
is solar energy. The correlated portion P0 is generated from a the dishwasher. The appliances’ turn ON probabilities are 0.01
solar irradiance data extracted from [19]. The data are of 1-min for both of them. Their energy demands are 1.772 kWh and
temporal resolution and was collected on May 10th, 2018. Each 1.186 kWh, respectively.
consumer is assumed to have 4 PV panels, each of area 1.5 m2 3) TCL Appliances: The measurements include two TCL
and of 20% power conversion efficiency. The uncorrelated por- appliances: a fridge and an air conditioning system. However, it
tion Piu is uniformly distributed of the same mean as P0 . The is difficult to deduce different TCL parameters from the power
weighting factor α is taken to be 0.9. measurements alone. Hence, we use TCL parameters from other
The data collected for the residential demand should be in an studies, as given in Table II.
appliance level with a temporal resolution of at least 1 min. The 4) Demand Diversity: The aforementioned extracted param-
energy required by an appliance when it is turned ON can usually eters reflect the energy consumption behavior of a single con-
be deduced or given in the appliance’s data sheet. However, sumer only. Due to the limited available measurements for other
the consumer’s behavior toward using this appliance cannot be consumers, we make use of the extracted parameters to repre-
easily deduced. The same appliance can be used differently sent the demand of a large number of consumers (1000). We
by different consumers. Therefore, a statistical, appliance-level diversify the demand by altering some of these parameters ran-
demand data should be collected by monitoring each appliance domly from a consumer to another. The alteration is uniformly
for a large number of consumers. Unfortunately, such data are distributed, and variations can reach up to ±10% of the original
not abundant, since the concern of the operators and planners of values. The altered parameters are the thermal resistance and
power networks is the aggregated demand of many consumers, capacitance of the TCL appliances, the energy consumption of
not the appliance-level demand. Nevertheless, there exists a the noncontrollable loads, and the switching probability of the
study [20] that recorded the appliance-level residential demand deferrable loads.
for a period of almost two years (2012–2014) for a small number Samples of an RES power, noncontrollable demand, and net
of consumers (five consumers only). Our approach is to use the noncontrollable demand profiles are shown in Fig. 3.
ELGHITANI AND EL-SAADANY: SMOOTHING NET LOAD DEMAND VARIATIONS USING RESIDENTIAL DEMAND MANAGEMENT 395

Fig. 3. Random samples for an RES power production of a single con-


Fig. 5. Total net demand profiles for different demand scheduling dead-
sumer (top), aggregated noncontrollable demand profile of all consumers
lines.
(middle), and the aggregated “net” noncontrollable demand profile of all
consumers (bottom), where large fluctuations are introduced by the ag- TABLE III
gregated RES contributions of all consumers. RELATIVE COST COMPARISON FOR THE PROFILES IN FIG. 5

Fig. 4. Decomposing total net demand profile into a smoothed profile


and an unmatched demand profile (Ψ = 60).

B. Performance Evaluation
We decompose the total aggregated demand profile into two
profiles: First, the demand supplied directly by the GENCO Fig. 6. Comparison between the proposed methodology and “looka-
and, second, the mismatch purchased from the AS market. For head” methodology.
simplicity, we calculate the first profile as the moving average of
the original profile with a window size same as the generator’s Theorem 1. A comparison of the relative cost for each of the
ramping time Ψ. Hence, the first profile represents a smoothed generated profiles is given in Table III.
version of the original profile. The mismatch profile is simply We compare our proposed methodology to “lookahead”
calculated by taking the difference between the original profile methodology used in several related works (see, e.g., [9]). In
and its smoothed version, as illustrated in Fig. 4. this methodology, RES profile is forecast first, then the control-
Several total demand profiles under our proposed methodol- lable demand is scheduled accordingly, as a solution of a deter-
ogy are depicted in Fig. 5. For simplicity, we assumed a single ministic convex optimization problem. The performance of this
class of controllable demands. From Fig. 5, it is clear that the methodology largely depends on the forecasting accuracy, be-
fluctuations are reduced as the demand scheduling deadline in- sides its excessive computational complexity. Our performance
creases. This behavior is intuitive since longer scheduling dead- measure is the relative cost of AS purchased, which is the ra-
lines will provide more flexibility for the demand controller. tio between the cost under DR and without DR (no delay). For
From a mathematical point of view, longer deadlines imply simplicity, we consider a single class of controllable demand.
larger values for the tradeoff parameter V , which allows the Fig. 6 plots the relative cost versus deadline for both method-
total net profile to approach its optimal shape, as implied from ologies assuming a 20% forecasting error for the latter. As the
396 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

following set of inequalities hold:


E{xi (t)}  E{ai (t)} ∀i ∈ Q. (14)
Proof: See [17, Appendix 4.A]. 
The importance of Lemma 1 is that it let us focus on ω-only
policies, i.e., policies that observe only the current state of the
system and not the entire system history. Thus, if we find a
policy that is optimal among ω-only policies, it will be optimal
among all other policies as well.
We proceed to use Lyapunov optimization that is supposed
to provide different limits on all queues (real and virtual),
in addition to minimizing the objective function in P1. Lya-
punov optimization avoids dynamic programming by develop-
ing a greedy algorithm (that will become Algorithm 1), which
minimizes both the objective function and the queueing cost.
Fig. 7. Performance under different number of queues normalized to
First, we introduce a measure of queueing congestion,
 referred

the performance of single queue. to as Lyapunov function L(Θ(t))  12 i wi Q2i (t) + Zi2 (t) ,
where Θ(t) is the state of all queues {Qi (t), Zi (t)} and {wi }
deadline increases, the performance of our methodology im- are proper weights to treat queues differently according to their
proves. In contrast, the performance of the “lookahead” method- classes. The queueing cost we aim to minimize is the conditional
ology deteriorates when the lookahead horizon (deadline) is Lyapunov drift given by
large. Δ(Θ(t))  E{L(Θ(t + 1)) − L(Θ(t))|Θ(t)}. (15)
Finally, we investigate the benefit of using a multiclass queue-
ing system. Increasing the number of queues allows more ef- The following lemma provides an upper bound on Δ(Θ(t)).
ficient assignment of controllable demand requests; however, Lemma 2: Assuming that scheduling decisions are bounded,
it decreases the arrival rate per queue, which in turn decreases i.e., 0  xi (t)  xm
i
ax
, then Lyapunov drift at any time slot t is
queues’ capability for regulation. Hence, we can expect that for upper bounded by the following expression:
each distribution of demand requests, there is an optimal num- N
 
ber of queues that achieves the best performance (lower cost). Δ(Θ(t))  B + wi Qi (t)E{ai (t) − xi (t)}
The cost of purchasing AS for different numbers of queues i=1
(normalized to the cost under single queue) is shown in Fig. 7. 
+ Zi (t)E{i − xi (t)}|Θ(t) . (16)
The figure shows that the performance improves rapidly as the
number of classes’ increases until the number of classes reaches Proof: From (8)
seven, after which the cost starts to increase slowly.
Zi (t + 1)  [Zi (t) − xi (t) + i ]+
V. CONCLUSION and hence
In this paper, we propose a methodology for residential de- wi Zi2 (t + 1)  wi [Zi (t) − xi (t) + i ]2 .
mand management to reduce the fluctuations imposed by a large-
scale integration of renewable energy sources. Our methodology Thus,
is composed of two steps: The first step is to classify the demand
wi [Zi2 (t + 1) − Zi2 (t)]
requirements of different residential appliances to fixed classes
of demands, according to their delay tolerances. The second  wi [i − xi (t)]2 + 2wi Zi (t)[i − xi (t)]
step is to use Lyapunov optimization to schedule the demands  
deferred in different queues. The resulting algorithm is compu-  wi max 2i , (xm i
ax 2
) + 2wi Zi (t)[i − xi (t)].
tationally efficient that makes it suitable for running in real time. Similarly from (7)
The methodology does not need highly restrictive assumptions
and it is proved effective under realistic data. wi [Q2i (t + 1) − Q2i (t)]
 wi [x2i (t) + a2i (t) − 2xi (t)ai (t)] + 2wi Qi (t)[ai (t) − xi (t)]
APPENDIX A
PROOF OF THEOREM 1  wi [(xm
i
ax 2
) + (am
i
ax 2
) ] + 2wi Qi (t)[ai (t) − xi (t)].
We begin with characterizing c∗ using the following lemma. Substituting these bounds in the definition of Lyapunov drift
Lemma 1: Assuming that the optimization problem (P1) is [see (15)] proves the lemma. 
feasible and that the system inputs, ω(t)  {ai (t), r(t)}, are The greedy algorithm is simply composed of minimizing both
independent and identically distributed (i.i.d.) processes, then the drift bound defined in (16) and the second moment of the
there exist an ω-only policy that can achieve c∗ . In addition, the total power consumption weighted by a tradeoff parameter V .
ELGHITANI AND EL-SAADANY: SMOOTHING NET LOAD DEMAND VARIATIONS USING RESIDENTIAL DEMAND MANAGEMENT 397

Thus, for each time slot t, we solve the following problem: By taking the expectation on both sides

2 ⎫
⎨ N
 ⎬ E{L(Θ(t + 1)} − E{L(Θ(t)}
P3 : min V E r(t) + xi (t) ⎧
2 ⎫
x i ,∀i∈Q ⎩ ⎭ ⎨ N
 ⎬
i=1
+VE r(t) + xi (t)  B + V c∗ . (19)
N
 ⎩ ⎭
i=1
− wi [Qi (t) + Zi (t)] E{xi (t)}. (17)
i=1 By summing from t = 0 to t = T − 1, then
Problem P3 can simply be solved by setting the gradi- E{L(Θ(T )} − E{L(Θ(0)}
ent of the objective function to zero, which yields (10) of ⎧
2 ⎫
Algorithm 1. T
 ⎨ N
 ⎬
We now focus on proving the part 1 of the theorem. Since +V E r(t) + xi (t)  BT + V T c∗ . (20)
⎩ ⎭
t=0 i=1
all queues were initially empty, then Qi (0)  Qm i
ax
. Assume,
m ax
at an arbitrary time t0 , that Qi (t0 )  Qi . If we attempt to However, L(Θ(0) = 0 (since queues are initially empty) and
prove that Qi (t0 + 1)  Qm i
ax
, then Qi (t)  Qm i
ax
for all t L(Θ(t)  0 for any time slot, and hence
by induction. We divide our solution into two complementary ⎧
2 ⎫
cases: First, when Qi (t0 )  Qm ax
− am ax
and, second, when T
 −1 ⎨ N ⎬
i i
Qim ax
− aim ax m ax
< Qi (t0 )  Qi . For the first case, it is obvious V E r(t) + xi (t)  BT + V T c∗ . (21)
⎩ ⎭
that Qi (t0 + 1)  Qm i
ax
, since Qi (t) can be incremented by at t=0 i=1
m ax
most ai . For the second case Dividing by V T completes the proof.
wi Qi (t0 ) > 2N V {N am
i
ax
+r m ax
}
APPENDIX B
and, hence, by using the result in (10) PROOF OF THEOREM 2
N
wi [Qi (t0 ) + Zi (t0 )] r(t0 ) We define the T -slot sample-path drift as follows:
xi = i=1 −
2N 2 V N
ΔT (Θ(t))  L(Θ(t + T )) − L(Θ(t))
wi Qi (t0 ) r(t0 )
 − > am i
ax
.
2N 2 V N which can be rewritten as follows:
We then divide the second case into two complementary sub- t+T
−1
cases: First, when Qi (t0 )  xi (t0 ), and second, when Qi (t0 ) > ΔT (Θ(t)) = L(Θ(t + j)) − L(Θ(t + j − 1)).
xi (t0 ). For the first subcase Qi (t0 + 1) = ai (t0 + 1)  am
i
ax
 j =t
m ax
Qi . For the second subcase, Qi (t0 ) will be decremented by From the proof provided in lemma 2
at least am i
ax
− ai (t0 + 1), and hence, Qi (t0 + 1)  Qi (t0 )  
t+T
−1 N
Qi . Hence, under all circumstances, Qi (t0 + 1)  Qm
m ax
i
ax
. 
m ax ΔT (Θ(t))  B+ wi Qi (j){ai (j) − xi (j)}
Similar argument can be made for Zi .
j =t i=1
The second part of the theorem can be proved as follows. 
According to the i.i.d. assumptions, the results in Lemma 1 can 
be applied. We refer to the decisions of the optimal ω-only + Zi (j){i − xi (j)}
policy as xω i . Our greedy algorithm is not an ω-only policy, 
as the scheduling decisions depend on Θ(t) as well; however, t+T
−1 N
 
it minimizes the objective function in P3 among all policies = BT + wi Qi (j){ai (j) − xi (j)}
including the optimal ω-only policy, thus j =t i=1

2 ⎫

⎨ N
 ⎬ 
Δ(Θ(t)) + V E r(t) + xi (t) + Zi (j){i − xi (j)} . (22)
⎩ ⎭
i= 1

2 ⎫ We attempt to provide an upper bound for the future values
⎨ N
 ⎬
B+VE r(t) + xω of various random variables. Since the maximum increment for
⎩ i (t) ⎭
i= 1 any queue is its maximum arrival rate, then
N
 Qi (j){ai (j) − xi (j)}  {Qi (t) + (j − t)am
i
ax
}{ai (j) − xi (j)}
+ wi [Qi (t)E{ai (t) − xω ω
i (t)} + Zi (t)E{i − xi (t)}|Θ(t)] .
i= 1
 Qi (t){ai (j) − xi (j)} + (j − t)(am
i
ax 2
)

By applying Lemma 2, it results and



2 ⎫
⎨ N ⎬ Zi (j){i − xi (j)}  {Zi (t) + (j − t)i }{i − xi (j)}
Δ(Θ(t)) + V E r(t) + xi (t)  B + V c∗ . (18)
⎩ ⎭  Zi (t){i − xi (j)} + (j − t)2i .
i=1
398 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

By substituting into (22) [13] J. Widén and E. Wäckelgård, “A high-resolution stochastic model of
domestic activity patterns and electricity demand,” Appl. Energy, vol. 87,
N
 t+T
−1
(2i + (am
i
ax 2
) )T (T − 1) no. 6, pp. 1880–1892, Jun. 2010.
ΔT (Θ(t))  BT + wi + [14] F. Elghitani and W. Zhuang, “Aggregating a large number of residential
i
2 j =t
appliances for demand response applications,” IEEE Trans. Smart Grid,
to be published.
N  [15] P. Du and N. Lu, “Appliance commitment for household load scheduling,”
 IEEE Trans. Smart Grid, vol. 2, no. 2, pp. 411–419, May 2011.
× wi [Qi (t){ai (j) − xi (j)} + Zi (t){i − xi (j)}] . [16] M. J. Neely, “Energy optimal control for time-varying wireless networks,”
i=1 IEEE Trans. Inf. Theory, vol. 52, no. 7, pp. 2915–2934, Jul. 2006.
(23) [17] M. J. Neely, Stochastic Network Optimization With Application to Com-
munication and Queueing Systems(Synthesis Lectures on Communica-
However, from the definition of B given in Theorem 1 tion Networks), vol. 3. San Rafael, CA, USA: Morgan & Claypool, 2010,
pp. 1–211.
N
 [18] M. J. Neely, A. S. Tehrani, and A. G. Dimakis, “Efficient algorithms for
(2i + (ami
ax 2
) )
wi B renewable energy allocation to delay tolerant consumers,” in Proc. 1st
i
2 IEEE Int. Conf. Smart Grid Commun., Oct. 2010, pp. 549–554.
[19] Website of National Renewable Energy Laboratory, Baseline
then Measurement System (BMS), May 2018. [Online]. Available:
http://http://www.nrel.gov/
t+T
−1 [20] J. Kelly and W. Knottenbelt, “The UK-DALE dataset, domestic appliance-
ΔT (Θ(t))  BT 2 + level electricity demand and whole-house demand from five UK homes,”
Sci. Data, vol. 2, Mar. 2015, Art. no. 150007.
j =t [21] E. Kara, M. Berges, and G. Hug, “Impact of disturbances on modeling
N  of thermostatically controlled loads for demand response,” IEEE Trans.
 Smart Grid, vol. 6, no. 5, pp. 2560–2568, Sep. 2015.
× wi [Qi (t){ai (j) − xi (j)} + Zi (t){i − xi (j)}] . [22] H. Hao, B. M. Sanandaji, K. Poolla, and T. L. Vincent, “A generalized bat-
i=1 tery model of a collection of thermostatically controlled loads for provid-
(24) ing ancillary service,” in Proc. 51st IEEE Annu. Allerton Conf. Commun.,
Control, Comput., Oct. 2013, pp. 551–558.
Starting from inequality (24) and after adding the cost term
j =t+T −1   2
V j =t r(t) + N i=1 xi (t) , we can follow the same
procedure used for proving theorem 1. The rest of the proof is
eliminated for brevity.

REFERENCES
[1] J. von Appen, M. Braun, T. Stetz, K. Diwold, and D. Geibel, “Time in the Fadi Elghitani received the B.Sc. and M.Sc. de-
sun: The challenge of high PV penetration in the German electric grid,” grees from Ain Shams University, Cairo, Egypt,
IEEE Power Energy Mag., vol. 11, no. 2, pp. 55–64, Mar./Apr. 2013. and the Ph.D. degree from the University of
[2] M. Liserre, T. Sauter, and J. Y. Hung, “Future energy systems: Integrating Waterloo, Waterloo, ON, Canada.
renewable energy sources into the smart power grid through industrial He is currently an Assistant Lecturer with
electronics,” IEEE Ind. Electron. Mag., vol. 4, no. 1, pp. 18–37, Mar. 2010. the National Telecommunication Institute, Cairo,
[3] F. Katiraei and J. R. Aguero, “Solar PV integration challenges,” IEEE Egypt. His research interests include smart grid
Power Energy Mag., vol. 9, no. 3, pp. 62–71, May/Jun. 2011. and wireless communications.
[4] E. Ela and M. O’Malley, “Studying the variability and uncertainty impacts
of variable generation at multiple timescales,” IEEE Trans. Power Syst.,
vol. 27, no. 3, pp. 1324–1333, Aug. 2012.
[5] N. Navid and G. Rosenwald, “Market solutions for managing ramp flexi-
bility with high penetration of renewable resource,” IEEE Trans. Sustain.
Energy, vol. 3, no. 4, pp. 784–790, Oct. 2012.
[6] E. Ela et al., “Evolution of operating reserve determination in wind power
integration studies,” in Proc. IEEE Power Energy Soc. Gen. Meeting, 2010,
pp. 1–8.
[7] E. Lannoye, D. Flynn, and M. O’Malley, “Evaluation of power system flex-
ibility,” IEEE Trans. Power Syst., vol. 27, no. 2, pp. 922–931, May 2012.
[8] B. Zakeri and S. Syri, “Electrical energy storage systems: A comparative Ehab El-Saadany (F’18) was born in Cairo,
life cycle cost analysis,” Renewable Sustain. Energy Rev., vol. 42, pp. 569– Egypt, in 1964. He received the B.Sc. and
596, Feb. 2015. M.Sc. degrees in electrical engineering from Ain
[9] Z. Tan, P. Yang, and A. Nehorai, “An optimal and distributed demand Shams University, Cairo, Egypt, in 1986 and
response strategy with electric vehicles in the smart grid,” IEEE Trans. 1990, respectively, and the Ph.D. degree in elec-
Smart Grid, vol. 5, no. 2, pp. 861–869, Mar. 2014. trical engineering from the University of Water-
[10] G. Wang, J. Zhao, F. Wen, Y. Xue, and G. Ledwich, “Dispatch strategy of loo, Waterloo, ON, Canada, in 1998.
PHEVs to mitigate selected patterns of seasonally varying outputs from He is a Professor with the Department of Elec-
renewable generation,” IEEE Trans. Smart Grid, vol. 6, no. 2, pp. 627–639, trical and Computer Engineering, University of
Mar. 2015. Waterloo. His research interests include smart
[11] P. Yang, P. Chavali, E. Gilboa, and A. Nehorai, “Parallel load schedule grid operation and control, power quality, dis-
optimization with renewable distributed generators in smart grids,” IEEE tributed generation, power electronics, digital signal processing applica-
Trans. Smart Grid, vol. 4, no. 3, pp. 1431–1441, Aug. 2013. tions to power systems, and mechatronics.
[12] Z. Zhu, S. Lambotharan, W. H. Chin, and Z. Fan, “A game theoretic op- Prof. El-Saadany is a Canada Research Chair in Smart Distribution
timization framework for home demand management incorporating local Systems, an Editor for the IEEE TRANSACTIONS ON SMART GRID and the
energy resources,” IEEE Trans. Ind. Inform., vol. 11, no. 2, pp. 353–362, IEEE TRANSACTIONS ON POWER SYSTEMS, and a Registered Professional
Apr. 2015. Engineer in the Province of Ontario.

You might also like