Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
ScienceDirect
AvailableScienceDirect
Available onlineatatwww.sciencedirect.com
online www.sciencedirect.com
Energy Procedia 00 (2017) 000–000

ScienceDirect
Energy Procedia 00 (2017) 000–000
ScienceDirect
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

Energy
EnergyProcedia
Procedia141 (2017) 000–000
00 (2017) 189–193
www.elsevier.com/locate/procedia

4th International Conference on Power and Energy Systems Engineering, CPESE 2017, 25-29
4th International Conference September
on Power and Energy
2017, Systems
Berlin, Engineering, CPESE 2017, 25-29
Germany
September 2017, Berlin, Germany
Utilization
The of
15thMLP and Linear
International Regression
Symposium Methods
on District Heating to Build a
and Cooling
Utilization of MLP and Linear Regression Methods to Build a
Reliable Energy Baseline for Self-benchmarking Evaluation
Assessing
Reliable the feasibility
Energy Baseline for of using the heat demand-outdoor
Self-benchmarking Evaluation
temperature function for a long-term
Tomorn district heat demand forecast
Sunthornnapha*
Tomorn Sunthornnapha*
Department of Electrical Engineering, Siam University, 38 Petchkasem Rd. Bangwa Phasrichareon, Bangkok 10160, Thailand
a,b,c a a b c c
I. Andrić *, A. Pina , P. Ferrão , J. Fournier ., B. Lacarrière , O. Le Corre
Department of Electrical Engineering, Siam University, 38 Petchkasem Rd. Bangwa Phasrichareon, Bangkok 10160, Thailand
a
IN+ Center for Innovation, Technology and Policy Research - Instituto Superior Técnico, Av. Rovisco Pais 1, 1049-001 Lisbon, Portugal
b
Abstract Veolia Recherche & Innovation, 291 Avenue Dreyfous Daniel, 78520 Limay, France
c
Abstract Département Systèmes Énergétiques et Environnement - IMT Atlantique, 4 rue Alfred Kastler, 44300 Nantes, France
This paper presents a reliable energy baseline model for self-benchmarking evaluation of energy saving potential by using
multilayer
This paperperceptron
presents a(MLP) method.
reliable energyThe measured
baseline modelenergy
for data and product quantities
self-benchmarking evaluationof the
of sample
energy plant
savingin daily period
potential by dating
using
back since perceptron
multilayer 2011 to 2016 are used
(MLP) as variables
method. and then
The measured normalized
energy data andtoproduct
represent the energy
quantities of thebaseline
sample(EnB)
plant inof daily
the manufacturing
period dating
Abstract
plant. A comparison
back since 2011 to 2016 of MLP andaslinear
are used regression
variables and then(LR) methods toforrepresent
normalized creatingthe theenergy
baseline model(EnB)
baseline is investigated during the
of the manufacturing
factory
plant. Aexpansion
comparison capacity.
of MLP Forand
LR linear
method, we use the
regression ASHRAE
(LR) methodsGuideline 14-2002
for creating as a reference
the baseline modelinisrecommended values the
investigated during for
District heating networks are commonly addressed in the literature as one of the most effective solutions for decreasing the
modeling uncertainty.
factory expansion As theFor
capacity. uncertainty
LR method, problem,
we usethetheLR method isGuideline
ASHRAE more sensitivity
14-2002toasthe outliners, in
a reference because the nature
recommended of plant
values for
greenhouse gas emissions from the building sector. These systems require high investments which are returned through the heat
variables
modeling has more complexity
uncertainty. and nonlinearity.
As the uncertainty problem,Sothe
we LRintroduce
methodthe is MLP
more method
sensitivityto solve
to theoroutliners,
reduce the effect the
because of nonlinearity by
nature of plant
sales. Due to the changed climate conditions and building renovation policies, heat demand in the future could decrease,
supervised
variables haslearning in the short-term
more complexity and long-term
and nonlinearity. period
So we of thethe
introduce production.
MLP methodFor simulation
to solve or results, in short-term
reduce the period the LR
effect of nonlinearity by
prolonging the investment return period.
method
superviseddemonstrates
learning in some better results
the short-term of uncertainty
and long-term periodparameters. However,For
of the production. thesimulation
proposed results,
MLP with LR method
in short-term can build
period the LRa
The main scope of this paper is to assess the feasibility of using the heat demand – outdoor temperature function for heat demand
reliable baseline showing
method demonstrates someinbetter
betterresults
R-square values thanparameters.
of uncertainty LR method. This is the
However, useful for energy
proposed MLP evaluation
with LR method when the
can plant
build isa
forecast. The district of Alvalade, located in Lisbon (Portugal), was used as a case study. The district is consisted of 665
expanding capacity
reliable baseline to protect
showing misleading
in better interpretation
R-square values than occurring
LR method. duringThis
the isyear. Forfor
useful long-term
energy period,
evaluation the when
MLP method
the plantcanis
buildings that vary in both construction period and typology. Three weather scenarios (low, medium, high) and three district
overcome
expanding the LR method
capacity in all
to protect uncertainty
misleading parameters. occurring
interpretation Therefore,during
the MLP method
the year. Formay be ableperiod,
long-term to the thealternative choice can
MLP method for
renovation scenarios were developed (shallow, intermediate, deep). To estimate the error, obtained heat demand values were
creating
overcomethetheEnB LRinmethod
nonlinearity
in all circumstances of the plants
uncertainty parameters. for short-term
Therefore, and long-term
the MLP method may period.
be able to the alternative choice for
compared with results from a dynamic heat demand model, previously developed and validated by the authors.
© 2017 The
creating Authors.
the EnB Published by
in nonlinearity Elsevier Ltd.of the plants for short-term and long-term period.
circumstances
©The
2017results showed that
The Authors.
Peer-review
when only
Published
under responsibility by weatherLtd.
of Elsevier
change is considered, the margin of error could be acceptable for some applications
the organizing
© 2017 The Authors. Published by Elsevier Ltd. committee of CPESE 2017.
Peer-review
(the error inunder responsibility
annual demand was of lower
the scientific
than 20%committee of the 4th
for all weather International
scenarios Conference
considered). on Power
However, afterand Energy renovation
introducing
Peer-review
Systems under responsibility of the organizing committee of CPESE 2017.
Engineering.
scenarios, the error value increased up to 59.5% (depending on the weather and
Keywords: energy conservation; uncertainty; energy baseline; multilayer perceptron; linear regression
renovation scenarios combination considered).
The value of slope coefficient increased on average within the range of 3.8%
Keywords: energy conservation; uncertainty; energy baseline; multilayer perceptron; linear regression
up to 8% per decade, that corresponds to the
decrease in the number of heating hours of 22-139h during the heating season (depending on the combination of weather and
renovation scenarios considered). On the other hand, function intercept increased for 7.8-12.7% per decade (depending on the
coupled scenarios). The values suggested could be used to modify the function parameters for the scenarios considered, and
improve the accuracy of heat demand estimations.

© 2017 The Authors. Published by Elsevier Ltd.

Peer-review under responsibility of the Scientific Committee of The 15th International Symposium on District Heating and
* Corresponding author. Tel.: +662-457-0068; fax: +662-457-3982.
Cooling.
* E-mail tomorn.soo@siam.edu
address:author.
Corresponding Tel.: +662-457-0068; fax: +662-457-3982.
E-mail address: tomorn.soo@siam.edu
Keywords: Heat demand; Forecast; Climate change
1876-6102 © 2017 The Authors. Published by Elsevier Ltd.
Peer-review
1876-6102 ©under
2017responsibility
The Authors. of the organizing
Published committee
by Elsevier Ltd. of CPESE 2017.
Peer-review under responsibility of the organizing committee of CPESE 2017.

1876-6102 © 2017 The Authors. Published by Elsevier Ltd.

Peer-review under responsibility of the Scientific Committee of The 15th International Symposium on District Heating and Cooling.
1876-6102 © 2017 The Authors. Published by Elsevier Ltd.
Peer-review under responsibility of the scientific committee of the 4th International Conference on Power and Energy
Systems Engineering.
10.1016/j.egypro.2017.11.036
190 Tomorn Sunthornnapha / Energy Procedia 141 (2017) 189–193
T. Sunthornnapha / Energy Procedia 00 (2017) 000–000

1. Introduction

Energy conservation in manufacturing plant is the energy-awareness management of modern industries. In many
countries, the factories have to be controlled and operated under the energy conservation act or energy conservation
standard such as ISO 50001:2011 [1]. The major procedure in this standard is to propose the energy saving measures
to minimize the energy consumption by setting priority on the maximum load to be done first.
Furthermore, ISO 50006:2014 standard [2] indicates that we must create energy baseline (EnB) and choose a
suitable energy performance index (EnPI) for comparison the saving between pre-retrofit and post-retrofit. Usually,
we use the specific energy consumption index (SEC) to determine the energy efficiency. In addition, the baseline
duration can range from less than one year to multiple years based on business condition information together with
statistical method such as regression analysis.
In energy saving potential, energy service companies (ESCOs) need to evaluate the feasible energy saving of the
factories or buildings by using the measurement and verification (M&V) procedures [3] and ASHRAE Guideline 14-
2002 [4]. The limitation of ASHRAE Guideline 14-2002 is the use of least square regression in approximation of
energy baseline and seems to be sensitive to outliners and nonlinear relationships.
Mostly reviewed papers, the research tasks are done in short-term (duration ≤ 12 months). There is an attempt as
in [5] proposed a comparison of Gaussian process regression and change-point regression for the baseline model in
industrial facilities. In uncertainty analysis, the application of Gaussian mixture models [6] is used to model the
baseline and localize adaptive uncertainty quantification. For weather season changing, there is a study of baseline
model using linear regression for office building energy consumption in hot summer and cold winter region [7] by
using monthly energy-bill. The application of multilayer feedforward artificial neural network is used for short
duration [8] with 15-minute aggregate energy data and shows the performance with more accurate results than a
baseline thin plate spline model.
At this point, the confusion of the duration used and the amount of information data for creating the baseline is
arisen on how we could apply these to the suitable method. So, we propose the MLP approach for approximation on
the reliable baseline, which rarely use in short-term and long-term analysis. Then, we can use this reliable baseline to
develop the EnPI for self-benchmarking evaluation for the factory that hardly to find the companies to benchmark
with or they have a big difference in technologies used in the process or machine.

2. Energy baseline approaches

In the case study, Mahasawat water treatment plant located in Nontaburi province, Thailand is chosen for
investigation the baseline. The energy consumption data of the plant equipment and facilities is measured from local
recordable power meters and complied into a daily report.
We can formulate the input-output variables of the system as independent variable vector of DDTF (Daily
distribution and transmission flow) and dependent variable vector of DPEC (Daily plant energy consumption) for
linear regression approach. For MLP approach, we use DRF (Daily raw water flow) as the independent variable to
predict DPEC and DDTF as shown in Fig.1.

Input
Plant Process
DRF phase1
[ DRF_ph1 : 1 ] Output EnB EnPI
DRF phase2 DDTF Linear
[ DRF_ph2 : 1 ] [ DDTF : 1 ] Regression
SEC

DRF phase3 DPEC

[ DPEC : 1 ] MLP
[ DRF_ph3 : 1 ]

DRF phase4
[ DRF_ph4 : 1 ]

Fig. 1 Mahasawat water treatment plant input-output formulated system

 Tomorn Sunthornnapha / Energy Procedia 141 (2017) 189–193 191
T. Sunthornnapha / Energy Procedia 00 (2017) 000–000

2.1. Linear Regression Techniques

In ASHRAE Guideline 14-2002, denotes that a linear regression is used to build energy baseline within the
coefficient of variation of the root mean square error (CV-RMSE) and the normalized mean bias error (NMBE) as
shown in Table 1. The linear regression is a combination of independent variables multiplied by constant as shown
in equation (1).

Table 1 Recommended values of baseline model uncertainty from ASHRAE Guideline 14

Baseline model uncertainty Monthly Hourly
CV-RMSE 15% 30%
NMBE 5% 10%

E C  B1V1  B2V2  B3V3   (1)

Where, E is the estimated energy consumption or demand by least square method, C the constant value in [energy
units/day] or [demand units/month], B n the coefficient of product variable Vn in [energy units/product/day] or
[demand units/product/day], and Vn the product variable.

2.2. Multilayer Perceptron

Multilayer Perceptron (MLP) is a feedforward neural network for function approximation. In [9] it has been
shown that the MLP ranked best for energy consumption estimation for the Canadian manufacturing industries,
followed by radial basis function network and support vector machine. MLP consists of input layer, hidden layers,
and output layer. Node i, also called a neuron, in an MLP network is shown in Fig. 2. It includes a summer and a
nonlinear activation function g.

Fig. 2 A multilayer perceptron network

The output yi, i =1,2, of the MLP network becomes

 3 2  3 K
 (2)
yi g 2   w 2ji g
1 (n j )   j 
1
g 2   w 2ji g1 (  wkj1 xk   1j )   j2 
 j1   j 1 k 1 

From (2) we know the set of input xk within the input layer then pass it through neurons or nodes multiplied by
weights wk. Then, sum it with bias j and activate the neurons nj by activation function g1 in hidden layer and g2 in
output layer. In this paper we chose g1 as tansig and g2 as purelinear [10] and use Levenberg-Marquardt method for
training and gradient descent optimization for learning and estimating the weights and biases of the final outputs y1
and y2.
192 Tomorn Sunthornnapha / Energy Procedia 141 (2017) 189–193
T. Sunthornnapha / Energy Procedia 00 (2017) 000–000

3. Simulation results

We classified the experiments into short-term and long-term period to study the behavior of variables affecting
the baseline. In short-term period, the results show that the EnB estimated from LR all years are in range of the
ASHRAE Guideline 14-2002. The MLP training results and validation are shown in Table 2 and then we get a better
training for MLP2. The CV-RMSE and NMBE values of LR approach are better than the uncertainty values derived
from MLP method as shown in Fig.3 and Table 3. However, R2 results from MLP are shown a superior than LR.

Table 2 MLP training results and validation

Methods No. of No. of hidden No. of R2 Mean square
inputs layers and neurons outputs of training vs target & error
validation vs target and epoch
MLP1 [12,071] 5 neurons [22,071] Training: 0.97026 0.008570
epoch =3
Validation: 0.975501
MLP2 [12,071] 10 neurons [22,071] Training: 0.975920 0.006346
epoch =6
Validation: 982402

a b

Fig. 3 Example of Short-term EnB : (a) LR; (b) MLP

Table 3 Short-term modeling uncertainty values (used daily data)

Methods Year R2 CV-RMSE (%) NMBE (%)
Linear regression 2011 0.7161 5.5139 0.0027
(Energy baseline, y=ax+b) 2012 0.9253 4.0762 -0.0011
x: Actual-DDTF 2013 0.8800 5.9141 0.0002
y: Actual-DPEC 2014 0.3157 8.7786 0.0032
2015 0.7972 5.9513 -0.0003
2016 0.8727 5.1585 -0.0019

MLP2 (best result) 2011 0.9634 6.0324 -0.0257

(Energy baseline, y=ax+b) 2012 0.9950 4.5843 0.1294
x: Training-DDTF 2013 0.9688 6.2420 -0.0533
y: Training-DPEC 2014 0.9342 9.2554 0.6159
2015 0.9311 6.6499 0.1472
2016 0.9453 4.4450 0.0154
 Tomorn Sunthornnapha
T. Sunthornnapha / Energy 00
/ Energy Procedia Procedia
(2017) 141 (2017) 189–193
000–000 193

a b

DPEC x 100 MWh

Fig. 4 Long-term EnB : (a) LR; (b) MLP

Table 4 Long-term modeling uncertainty values (used daily data)

Methods Year R2 CV-RMSE (%) NMBE (%)
Linear regression 2011-2016 0.8726 7.2080 0.0039
MLP2 2011-2016 0.9775 6.7533 -0.0008

For long-term period, we use the data from 2011 to 2016 as a training data for MLP approach. The results shown
in Fig.4 and Table 4 clearly demonstrate that MLP is a superior in all parameters of uncertainty. If we see the result
in the year of 2014 from Table 3, R2 value is lowest as 0.3157 and NMBE value is largest as 0.0032. This is because
the plant has a big variation in distribution pressures and has structural changes in energy use. Hence, the MLP may
be the suitable method for this situation if we use the short-term or long-term model to create a reliable baseline.

4. Conclusion and discussion

MLP approach for approximation on the reliable energy baseline gives the better result of uncertainty than the LR
when considered by the ASHRAE Guideline 14-2002 in long-term period. CV-RMSE value of the MLP is less than
the LR’s by 6.31% and NMBE value of MLP is also less than the LR’s by 79.49%. In short-term, the R2 values of
the MLP data are shown a very significant relationship between DDTF and DPEC showing the reliable baseline.

References

[1] ISO 50001:2011, Energy management systems, Requirements with guidance for use; 2011.
[2] ISO 50006:2014, Energy management systems, Measuring energy performance using energy baselines (EnB) and energy performance
indicators (EnPI), General principles and guidance; 2014.
[3] M&V Guidelines: Measurement and Verification for Performance-Based Contracts Version 4.0, Prepared for the U.S. Department of Energy
Federal Energy Management Program; November 2015.
[4] ASHRAE Guideline 14: Measurement of Energy and Demand Savings, ASHRAE Inc, Atlanta, GA; 2002.
[5] Joseph Carpenter et al., A Comparison of Gaussian Process Regression and Change-point Regression for the Baseline Model in Industrial
Facilities, ASHRAE and IBPSA-USA SimBuild 2016 Building Performance Modeling Conference, Salt Lake City, UT; August 8-12, 2016
[6] Abhishek Srivastav, Ashutosh Tewari, Bing Dong, Baseline building energy modeling and localized uncertainty quantification using
Gaussian mixture models, Journal of Energy and Buildings 65; 2013, p.438 -447.
[7] Fei Lei , Pingfang Hu, A baseline model for office building energy consumption in hot summer and cold winter region, International
Conference on Management and Service Science, MASS '09; 2009.
[8] Andrew Scott, Nabil Nassif, Darrion Long, Forecasting Savings of Building Energy Systems using Artificial Neural Networks, FTC 2016 -
Future Technologies Conference 2016, San Francisco, United States; 6-7 December 2016.
[9] Oludolapo Akanni Olanrewaju, Charles Mbohwa, Comparison of Artificial Intelligence Techniques for Energy Consumption Estimation, 16th
annual IEEE Electrical Power and Energy Conference (EPEC), 2016, Ottawa, Canada; 12-14 October 2016.
[10] Mark Hudson Beale, Martin T. Hagan, Howard B. Demuth, Neural Network Toolbox™, User's Guide, The MathWorks, Inc.; 2017.