Applied Energy 327 (2022) 120102

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

A novel data-driven optimal chiller loading regulator based on backward

modeling approach
Kuang-Yow Lian *, Yong-Jie Hong , Che-Wei Chang , Yu-Wei Su
Department of Electrical Engineering, National Taipei University of Technology, No. 1, Section 3, Zhongxiao East Road, Taipei 106, Taiwan

H I G H L I G H T S

• Developed a novel optimal chiller loading regulator to reduce energy consumption in industries.
• Deep neural network and conditional generative adversarial network were employed to achieve energy saving.
• Practical feasibility is verified by conducting 1-year field validation in panel manufacturing factory.
• Energy can be saved in the range of 81.9 MWh to 198 MWh per year by employing OCLR.

A R T I C L E I N F O A B S T R A C T

Keywords: This paper proposes a new method termed backward modeling approach (BMA) to achieve optimal chiller
Optimal chiller loading loading (OCL) for reducing energy consumption in industries running multiple-chillers with different efficiency.
Generative adversarial network The developed OCL regulator (OCLR) based on novel BMA approach is composed of conditional generative
Deep neural network
network (cGAN) and deep neural network (DNN). Most works on the optimal chiller loading problem are to find
Energy saving
out the setting of partial load rate (PLR) for each chiller. However, PLR for each chiller cannot be controlled
directly and can only be achieved through setting chilled water supply temperature. A novel feedback control
framework was developed to identify the relationship between chilled water supply temperature and the PLR. In
light of this, the control instruction for chilled water supply temperature can be set to achieve the desired energy
saving. The practical feasibility of loading optimization based on developed OCLR was evaluated by conducting
field validation for 1 year in a reputed panel manufacturing factory based in Taiwan running multiple-chiller
system. From the experimental results, it is evident that the developed data-driven OCLR based on BMA has
very high performance and was able to conserve significant energy in the range of 81.9 MWh to 198 MWh per
year.

1. Introduction consumption in chiller systems [2,3]. Even though various conventional

methods were applied, saving energy especially in multiple-chiller sys­
Chiller systems maintain the thermal condition of factories or tems still remains as a challenge to be addressed. Each chiller in the
buildings, providing cooling and ventilation to keep the space multiple-chiller system has different capacity and performance feature
comfortable for occupants and smooth running of equipment. The making its optimal operation by consuming minimum energy under
chiller system consumes about 60 % of the energy used for air- different cooling loads an arduous task [4–6]. Optimal chiller load dis­
conditioning, which accounts about 25 % to 40 % of energy in any tribution (OCL) of the multiple-chiller system can reduce the energy
commercial or industrial building [1]. Moreover, in many semi­ consumption and is thus an important topic to be investigated [7–9].
conductor fabrication plants and panel manufacturing factories, the Algorithms such as Lagrangian method (LM) were used in the liter­
chiller system is one of the major equipment that consumes high energy. ature to solve the OCL problem in multiple-chiller system optimization
This huge energy needed to operate chiller system is a major concern and improve the pitfalls of conventional methods [10]. However, the
and thus extensive research was carried out to reduce energy common pitfall of these algorithms is its failure to converge at low

* Corresponding author.
E-mail address: [email protected] (K.-Y. Lian).

https://doi.org/10.1016/j.apenergy.2022.120102
Received 27 March 2022; Received in revised form 7 September 2022; Accepted 2 October 2022
Available online 12 October 2022
0306-2619/© 2022 Elsevier Ltd. All rights reserved.
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 1. The feedback control framework for multiple-chiller system.

Fig. 2. The sketch of the multiple chiller system considered in this wok.

demand. Also, it is not adapted in solving OCL problems as the power We notice that most approaches in the literature focused on setting of
consumption and kilowatt-partial load rate (PLR) may contain convex PLRi [20,21]. However, it is worth noting that PLR for each chiller
and nonconvex functions [11]. Nevertheless, different approaches were cannot be controlled directly and can only be achieved by tuning chilled
adopted by researchers to overcome this problem. For instance, Chang water supply temperature (Tchws ). In fact, we can illustrate the rela­
et al. proposed a simulated annealing approach to solve the OCL prob­ tionship between Tchws and PLR using the feedback control framework as
lem and a gradient method was used to solve the economic dispatch of shown in Fig. 1, where the controller built in the chiller will drive the
chiller plant (EDCP), resulting in high accuracy results within a rapid Tchws to track the desired chilled water supply temperature (T * chws ).
timeframe [11]. A combination of genetic algorithm (GA) and particle Then, Tchws combining other environmental parameters, condenser
swarm optimization (PSO) was used by Ardakani et al. to solve the OCL water return temperature (Tconr ) and cooling load change the value of
problem. This combination approach showcased a fast convergence, PLR. In practical applications, Tchws must be restricted to a certain range,
simple implementation, escape from local optima, and independency of which implies that many high values PLRi reported in the literature are
the solution from the problem. Moreover, Zheng et al. applied an onerous to be implemented in practical scenario [22,23]. In this work,
improved artificial fish swarm algorithm for tackling the OCL problem, we will focus on discussing the setting of the optimal T * chws for each
where the objective function is the minimal power consumption of chiller to achieve optimal chiller loading (OCL) for reducing energy
chillers and cooling towers [1]. Furthermore, Beghi et al. [12] used a consumption in industries running multiple-chillers with different
PSO based method and Lee et al. applied a differential evolution algo­ efficiency.
rithm in optimal multiple-chiller system operation to avoid OCL To this end, we propose a new method termed backward modeling
problem. approach (BMA) to derive the optimal T* chws . The developed OCL
In the recent years, deep neural network approach finds its appli­ regulator (OCLR) based on the novel BMA approach is composed of
cation in many complex systems as it boasts significant merits compared conditional generative adversarial network (cGAN) [24] and deep
to the conventional methods [13,14]. Moreover, the availability of huge neural network (DNN) [25]. For feature selection, Pearson product-
amount of data recently aids in development of better and more efficient moment correlation coefficient (PCCs) [26] was iteratively used to
building and environmental models [15–19]. In this scenario, incorpo­ calculate the correlation coefficients among environmental parameters
ration of deep learning methods is an apt approach for efficient optimal and chilled system parameters. The proposed OCLR in this work can be
chiller load distribution in multiple-chiller system to reduce energy considered as an efficient approach to achieve loading optimization for
consumption. multiple-chiller systems. As per the best knowledge, this is the first paper

2
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 3. Overall chilled water system regulated by OCLR: (Left) Chilled water system and the collected data, where the data marked in light color are not used in the
design of OCLR; (Right) The schematic diagram of OCLR implementation.

that uses a feedback control system to identify the relationship between to be satisfied:
chilled water supply temperature and PLR of multiple chiller system. ∑l
Fig. 2. shows the overall chilled water system regulated by OCLR i=1
PLRi • RT i ≥ CL (3)
considered in this paper, where the cooling water and chilled water of 4
where RT i denotes the chiller capacity. Since each chiller has an
chillers connected in parallel are circulated by several distribution
appropriate load range, it is consequential to require PLRi lying in the
pumps.
following interval:

2. Preliminary of OCLR PLRi,min ≤ PLRi ≤ PLRi,max (4)

where PLRi,min and PLRi,max denote the minimum and maximum limit
The design of OCLR involves the following steps: Primarily, the raw
of PLRi , respectively. The setting of PLRi,min and PLRi,max is critical. At
data of the chilled water system are collected through the data gateway.
first glance, we have the impression that the larger the range is given,
Those data are used to train two models: a cGAN model for determining
the more electric energy consumption the system is going to save.
PLRi and a DNN model for power consumption estimation. Then, find
However, in practical application this is not the case as the range of PLRi
the optimal T* chws,i by minimizing the energy consumption of all chillers. is dominated by the chilled water supply temperature. To the best of our
Here, the total output cooling capacity of all the chillers must satisfy the knowledge, only few literatures have discussed how to set the optimal
actual cooling load (CL) of the factory. Finally, return the T* chws,i to the T* chws . We will deal with this issue in Section 4.
chilled system through the data gateway to control the chiller to achieve
energy saving, as shown in Fig. 3.
2.2. Establishing power consumption model
2.1. Basics of OCL
The power consumption model Pch,i is a kernel part of OCL. Hence,
The purpose of OCL is to meet the needs of CL by consuming mini­ establishing a model with minimum error is critical as it will directly
mum power under the restricted conditions of chiller operation. The impact the efficiency of OCL. Traditionally, power consumption models
power consumption for the i-th chiller can be expressed as follows: are mostly developed based on cubic polynomial or second-order poly­
nomial [27]. In these approaches, regression analysis was utilized to
Pch,i = fi (PLRi ) (1)
determine the respective coefficients.
which means that the consuming power Pch,i is the function value fi In this work, we used DNN to build the power consumption model.
determined by the partial load ratio PLRi . The goal of loading optimi­ The input parameter selected in this paper includes PLRi only. The DNN
zation is to minimize the energy consumption of l chillers and can be architecture used in this paper has a total of 3 hidden layers, the first
expressed as follows: layer has 128 neurons, the second layer has 64 neurons, the third layer
∑l has 32 neurons, and the output is Pch,i . The activation function used in
argmin i=1
fi (PLRi ) (2) DNN is rectified linear unit (ReLU) [28] and mean square error (MSE) is
used as the loss function. According to the design based on the rules of
PLRi ,i=1,..l

Simultaneously, the total cooling capacity of all the chillers must experience and multiple real data validations, the architecture has a
satisfy the actual CL of the factory, which yields the following constraint good accuracy rate.

3
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 4. Forward modeling architecture for multiple-chiller system.

Fig. 5. (Left) Data distribution for Tchws,i on PLRi and enthalpy based on forward modeling approach; (Right) Data distribution for PLRi on enthalpy and Tchws,i based
on BMA.

Fig. 6. Optimal chiller loading regulator for multiple-chiller system.

3. BMA and cGAN model for partial load rate next focus. Ideally, we hope to establish a forward model from PLRi to
Tchws,i , whereby we can directly substitute the best PLR* into the forward
The OCL described in the previous section can obtain the best PLR* , model to obtain T* chws , as shown in Fig. 4.
whereas how to obtain the adjustable parameter Tchws,i from PLRi is the However, such a direct method is actually impossible to realize. First

4
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 7. Pearson correlation coefficients of PLRi and other parameters.

of all, the Tchws,i obtained in this way does not necessarily meet the actual 3.1. Feature selection
operating range of the chiller system. In the application encountered in
this research, the suitable range of Tchws,i is 13 ℃ and 15 ℃, and can even In order to achieve an accurate and robust model for PLRi , we need to
be extended to 11 ℃ and 16 ℃. However, the best PLRi obtained by OCL take all possible features into consideration. For feature selection, we
is often far beyond what Tchws,i can actually achieve. In addition, it is not used PCCs to calculate those parameters that have high correlation co­
easy to establish a forward model from PLRi to Tchws,i . Even if we take efficients with PLRi . PCCs for parameters x and y is described as follows:
into consideration the relatively close parameter enthalpy of outdoor air ∑n
i=1 (xi − x)(yi − y)
(En), it affects the conversion from PLRi to Tchws,i significantly. The r(x, y) = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n ̅ (5)
2 ∑n 2
operating points corresponding to different Tchws,i are shown in Fig. 5 i=1 (xi − x) i=1 (yi − y)

(Left). It is found that the operating points overlap each other so much
where x and y denote the mean of x and y, respectively. According to
and is almost impossible to be classified. This inspired us to solve the
the field data, PCCs are calculated and shown in Fig. 7. At first, we may
problem via a backward method. Regarding the model from Tchws,i to
use all these parameters with relatively large correlation coefficients as
PLRi , when we take into account the closely influential parameter En,
input features of the model. Nevertheless, after in-depth analysis we
the operating points corresponding to different PLRi will be as shown in
found that these parameters can be divided into two categories. One
Fig. 5 (Right). Compared with Fig. 5 (Left), it is found that the color
category includes the parameters affected by PLRi , such as primary
blocks are relatively separated. Although there are still some mixed
pump frequency, primary pump power, etc. The other category includes
color patches, we will introduce other features in the next section to
the parameters that affect PLRi , such as En, dry-bulb temperature (DBT),
improve the accuracy of classifying PLRi .
wet-bulb temperature (WBT), absolute humidity (AH), CL, Tconr,i , and
Therefore, the appropriate approach will be as follows: First, estab­
Tchws,i . Since our target was to predict PLRi using these parameters and
lish a model from Tchws,i and other parameters to estimate PLRi , i.e., the
Tchws,i , it was inevitable to exclude the input features affected by PLRi .
cGAN Model for PLR in Fig. 3; For all Tchws,i and other parameters that
Consequently, the features selected to estimate PLRi in this paper
meet the actual operating range of the chiller system, search for the
include En, DBT, WBT, AH, CL, Tconr,i , and Tchws,i , which will be abbre­
corresponding PLRi ; Input all possible PLRi to the DNN power con­
viated as θi , i.e.,
sumption model to find the corresponding value kWi . Then, loading
( )
optimizer is applied to find the optimal PLR*i ; Subsequently, find out the θi = En, DBT, WBT, AH, CL, Tconr,i , Tchws,i (6)
corresponding parameter value T * chws,i according to the previous step, as
shown in Fig. 6. We call this approach BMA (backward modeling
approach). In this study, we also consider the status of turn on or turn off 3.2. Generative adversarial network
for each chiller as a state, namely Statei . Hence the number of running
chillers is imposed in our minimizing electricity consumption problem. In practical chillers, the operating range of Tchws,i is often set
The core of the backward modelling approach is the conversion conservatively. Although air-conditioning systems in semiconductor
model from Tchws,i to PLRi . Its accuracy can be very high. Combining with plants or panel manufacturers consume significant amounts of power,
the power consumption model of PLRi , as shown in Fig. 6, we can most manufacturers set all Tchws,i to the same value in a narrow range to
directly search for the total power consumption corresponding to the achieve stable operation [29]. For instance, according to the current
possible Tchws,i distribution under all constraining conditions. Then, we status of the manufacturer, the suitable range of Tchws,i is 13 ℃ and 15 ℃
can find the optimal T* chws,i distribution. and can even be extended to 11 ℃ and 16 ℃. Because of the conser­
vative mentality of engineers, Tchws,i was actually set between 13.7 and

5
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 8. The photo of multiple-chiller system considered in this work taken at the panel manufacturing factory.

14.3 for a long time. In the range of only 0.6 ℃, OCL has limited effects. power consumption model to find the optimal value T* chws,i . In load
To improve the efficiency of energy saving, it is worthwhile to expand distribution, the practical and feasible PLRi needs to meet the allowable
this range. However, this leads to the problem of lacking historical data range that chillers can achieve. Let θi |Tchws,min and θi |Tchws,max denote the
for chillers operating beyond 13.7 ℃ to 14.3 ℃. To overcome this selected features by conditioning on Tchws,min = 13 ℃ and Tchws,max = 15
problem, this work employed cGAN [30,31] to establish a conversion ℃, respectively. According to the model G established by cGAN, we may
relationship between the selected features θi and PLRi . rewrite the Eq. (4) for PLRi to a feasible one in terms of allowable in­
The framework of cGAN is developed based on the min–max method terval of Tchws,i as follows:
[32]. The principle is to learn a generator G, which synthesizes samples ( ) ( )
similar to the data distribution, such that the discriminator cannot G θi |Tchws,max ≤ G(θi ) ≤ G θi |Tchws,min (8)
distinguish between synthetic samples and real samples. On the other
hand, the objective is also to learn a good discriminator D so that it can Simultaneously, the total output cooling capacity of all the running
distinguish between synthetic samples and real samples. chillers must satisfy the actual CL of the factory. We used Statei = 1 or
There is no control on modes of data generated in the case of un­ 0 to denote that the i-th chiller is turn-on or turn-off, respectively. Ac­
conditional GAN. On account of the limitation in training process of cording to Eq. (3), we rewrite it as:
GAN, the proposed cGAN algorithm enact a constraining and guiding ∑l
role by concurrently adding conditional variables to input of generator i=1
G(θi ) • RT i • Statei ≥ CL (9)
and discriminator. In this work, the conditional variable is the chilled
water supply temperature. Conditional variables can be viewed as the Therefore, the OCL problem is to find the optimal T* chws,i and State*i
labels in supervised learning. Therefore, we can directly use cGAN to for the objective function (2) under Eq. (8) and Eq. (9) as follows:
establish the mapping relationship between the chilled water supply ( ) ∑
l

temperature and PLRi . T * chws,i , State*i = arg min fi (G(θi ) • Statei ) subject to (8) and (9).
In this work, cGAN is developed to establish a model from the
Tchws,i , Statei , i=1,..l i=1

selected features to estimate PLRi . For the cGAN, G learns a nonlinear (10)
mapping to synthesize the output PLRi by conditioning on the selected Referring to the flowchart in Fig. 6, we summarize the steps to
features θi , whereas D needs to distinguish between PLRi estimated by G perform OCLR as follows:
and extracted from real data PLRi : Step 1 Practical and feasible PLRi : Use G(θi ) model to calculate PLRi
based on current environmental parameters θi with Tchws,i from 13 ℃ to
minmaxV(G, D) = log(1 − D(θi , G(θi ) ) ) + logD(θi , PLRi ) (7)
G D 15 ℃ at 0.1 intervals.
Step 2 Possible load distribution: Look for load distribution for
For our framework, G and D are composed of DNN. The generator G
multiple chillers that meet Eqs. (8) and (9).
has 4 hidden layers in total, where the first layer to the fourth layer has
Step 3 Possible power consumption: Calculate the power consump­
128, 64, 32, and 16 neurons, respectively. The discriminator D has 3
tion of each load distribution.
layers of hidden layers in total, where the first layer to the third layer has
Step 4 Loading optimization: Find the smallest energy consumption
64, 32, 16 neurons, respectively. The training of the cGAN model is more
value.
complicated than the previous DNN model. In order to avoid the use of
Step 5 Output result: Output optimal chilled water supply tempera­
ReLU, the neurons will be necrotic, so Leaky ReLU is used as the acti­
tures and chiller status.
vation function of cGAN. Due to the large difference in the order of the
numerical values, batch normalization is used after each hidden layer.
5. Model validation and experimental results
4. Implementation of OCL via T chws,i
We have carried out the experiment in a famous panel manufacturer
located in Longtan Science-based Park, Taiwan to validate the practical
The previous section established a model from the selected features
feasibility of the developed model. This multiple-chiller system has a
including Tchws,i to estimate PLRi . This section will describe how to find
total of 4 chillers, each chiller capacity is 1600 RT, as shown in Fig. 8.
the corresponding PLRi for all Tchws,i and other parameters that meet the
The raw data of the third chiller and the established models using
actual operating range of the chiller system. We then use the DNN based

6
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 9. The kW-PLR diagram for the third chiller.

Table 1 Table 2
The data of performance coefficients of chillers. Verification results of DNN based power consumption model and cGAN model
Chiller ai bi ci di (RTi )
for multiple chillers.
Models Power consumption model
1 − 179.205 1113.760 − 176.945 7.960 1600
2 − 268.048 1529.554 − 859.156 370.754 1600 Chiller 1 Chiller 2 Chiller 3 Chiller 4
3 − 175.577 1218.657 − 481.918 167.425 1600
4 − 87.918 827.384 104.017 − 136.307 1600 DNN MAE (W) 84 91 134 135
Regression MAE (W) 87 99 136 137

cGAN model for PLR estimation

Chiller 1 Chiller 2 Chiller 3 Chiller 4

MAPE (%) 1.15 0.46 0.76 1.2

DNN and typical regression method are shown in Fig. 9. At first glance,
the DNN model and the typical regression model appear very similar in
the full-scale diagram shown in Fig. 9 (d). However, after zooming in the
scales, one can find that the DNN model is closer to the distribution of
raw data as shown in Fig. 9 (a), (b), and (c). Eq. (11) denotes the
respective regression equation.

Pch,i = ai + bi PLRi + ci PLR2i + di PLR3i (11)

where ai , bi , ci , di denote coefficients of the kW-PLR curve of the i-th

chiller. Its coefficients are shown in Table 1.
The relationship between partial load rate and power consumption of
the four chillers are depicted in Fig. 10. We noticed that even though the
capacity and the type of all the four chillers are the same, their char­
acteristic curves are not identical. This is due to the long-term running of
Fig. 10. The relationship between partial load rate and power consumption of the chillers.
four chillers.

7
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 13. Power consumption curve for the whole year using OCLR (Tchws,i is 11
Fig. 11. Power consumption curve without considering Tchws,i .
to 16 ℃).

Fig. 12. Power consumption curve for the whole year using OCLR (Tchws,i is 13
℃ to 15 ℃).

Fig. 14. Percentage error - Tchws,i diagram of four chillers.

Table 3
Optimal chiller loading results.
than 100,000 records. The verification results of DNN based power
CL CH ECHWT OCLR
∑ ∑ consumption model and the cGAN model for PLR estimation are shown
T* chws,i PLRi fi (PLRi ) T* chws,i PLRi fi (PLRi )
in Table 2. The results depicted in Table 2 clearly show that DNN model
4034 1 14.1 60.8 1784 15.0 57.1 1776 perform better than the regression model for all the chillers. As to the
(RT) 2 14.1 63.2 (kW) 15.0 59.6 (kW) proposed cGAN model for PLR estimation, the average mean absolute
3 14.1 65.2 13.0 68.3
percentage error (MAPE) of multiple chillers is mere 0.89 %. From the
4 14.1 62.9 13.4 67.1
experimental results, it is evident that the operation and power con­
sumption model established using the big data-driven OCLR has very
3981 1 14.0 61.4 1755 15.0 57.4 1746
high performance.
(RT) 2 14.0 60.2 (kW) 15.0 56.4 (kW)
3 14.0 66.6 13.0 71.0
4 14.0 60.6 13.1 64.0
5.2. Numerical simulations
3571 1 14.0 56.1 1536 15.0 50.4 1526
(RT) 2 14.0 54.1 (kW) 14.8 51.1 (kW) We have also carried out a comparative evaluation of energy saving
3 14.0 58.8 13.3 61.8 capability of the proposed OCLR with Equal Chilled Water Temperature
4 14.0 54.2 13.0 59.9 (ECHWT) method in this work [29]. ECHWT is commonly used for
chiller loading distribution in many factories. The panel manufacturer
5.1. Model verification where we have collected the experimental data was also using a similar
approach. The efficiency of energy saving using both approaches was
We used 80 % of the multiple-chiller data acquired from the panel evaluated by considering the data for a period of 1 year and is described
manufacturing factory during the year 2018 for model training and 20 % in detail in the following subsections.
data for model validation. The data of multiple chillers in 2018 is more The OCLR proposed in this study is to optimize the chilled water
supply temperature of each chiller (Tchws,i ) under the condition that the

8
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 15. Operating curves for power consumption, RT, DBT, and WBT for using ECHWD and OCLR on July 21 and July 24, 2020, respectively.

Fig. 16. Operating curves of PLR for using ECHWD and OCLR on July 21 and July 24, 2020, respectively.

cooling load or the overall averaged supply temperature of chilled water causing their effects on the power consumption of primary pumps to
(Tchws ) remain unchanged. That is, we can find the most power-saving cancel each other out. Hence, the deviation for the total power con­
distribution of PLRi while the cooling load remains the same. There­ sumption of primary pumps due to the change of distribution of PLRi is
fore, the setting of this problem hardly affects the power consumption of very small and can be negligible.
condenser pumps, cooling towers and secondary pumps. As for each
primary pump, its operating frequency is positively correlated with 5.2.1. Realization without Considering T chws,i
PLRi . We have observed carefully and found that the frequency of the Regardless of the allowable variation range of temperature Tchws,i , the
primary pump changes slightly even when the change of PLRi is large. PLRi range is set between 50 % and 80 %. Primarily, we conducted
Moreover, when OCLR is performed, the total PLR hardly changes. An energy saving analysis for OCL. Fig. 11 shows the annual power con­
increase in one PLRi will inevitably lead to a decrease in other PLRi , sumption of the chiller system and the power consumption after using

9
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Table 4 5.2.2. Practical realization with feasible T chws,i

Average data before and after using OCLR. According to the actual application scenario, we require Tchws,i
Parameters ECHWT (A) OCLR (B) Deviations (A - B) varying between 13 ℃ and 15 ℃. Fig. 12 shows the overall annual
power consumption improvement after the adoption of OCLR.
Time July 21, 2020 July 24, 2020 (288 samples totally)
Averageconsumption 1975.12 2009.72 − 34.6 Compared to the realization without considering Tchws,i as mentioned in
(kW) sub-section 5.2.1, in this case, the overall energy saving effect is some­
Averagecooling load 4372.85 4493.84 − 121 what lower, especially from April to October. The reason is that when
(RT)
the weather is hot, relatively high CL and En lead to a decrease in the
Average DBT (˚C) 28.25 29.71 − 1.46
Average WBT (˚C) 23.92 24.56 − 0.64
ability of Tchws,i to affect PLRi . In our application, although the charac­
Averageefficiency 0.45168 0.44722 0.00446 (0.99 %) teristics of the four chillers are different, the insufficient difference leads
(kW/RT) to limited energy saving efficiency.
According to the numerical evaluation, the annual electric energy
consumption of OCLR is about 14,807 MWh, which is a reduction of 81.9
MWh (save about 0.55 %) for the whole year compared to the practical
annual electric energy consumption of 14,888.9 MWh. In Table 2, the
MAE of chiller power consumption models is 0.8 ~ 1.4 kW. This means
that the average error of chiller power consumption models with actual
chiller power consumption is about 1.1 kW, which is a very small error.
Therefore, the 81.9 MWh (save about 0.55 %) saved by OCLR is reliable.
We selected several energy saving improvement cases, and listed their
detailed values for the loading distribution results in Table 3 for refer­
ence. From the load distribution resulting from OCLR, it can be found
that chiller 1 has the lowest load distribution indicating its poor
efficiency.
Furthermore, we were interested in exploring how much energy
saving can be further achieved by relaxing the range of Tchws,i . If the
allowable variation range of Tchws,i is set between 11 ℃ and 16 ℃, it is
found that the energy saving efficiency will be greatly improved to 1.33
%, as shown in Fig. 13. Compared with the previous design of Tchws,i
between 13 ℃ and 15 ℃, an additional increase of 0.78 % is obtained. If
Tchws,i is allowed to vary in a wider range, it can be expected that the
energy saving efficiency will continue to increase.
Fig. 17. The variation of PLRi , where the ECHWT is switched to OCLR at 13:30.

5.3. Practical feasibility of proposed cGAN model

Generally, the data-driven neural network models will have reliable

accuracy only if sufficient and complete data is available to train the
model. This is the reason why we introduced cGAN to build the model.
Here, we designed this experiment to verify the accuracy of the model
when Tchws,i is not between 13.7 ℃ and 14.3 ℃. The experimental results
are shown in Fig. 14. The four curves represent four different chillers,
where the horizontal axis denotes Tchws,i , and the vertical axis is the
MAPE of PLRi for the actual measured value and the estimated value
using the OCLR model. The experiment was carried out to adjust Tchws,i
for the four chillers from 13 ℃ to 15 ℃. We used 5 min as a unit and
recorded data 6 times for each chiller. From Fig. 14, we can find that
although the amount of data is insufficient, we can still accurately es­
timate the PLRi , and the error is within 1.6 %.

5.4. Field experiment for energy saving

We have carried out field experiment in a panel manufacturing fac­

Fig. 18. Power consumption curves for ECHWT and OCLR. tory and the proposed OCLR method has been applied to the multiple
chillers operating in the factory. In this study, the OCLR program was
OCL. We used brute force search (BF) to find the best load distribution incorporated to the monitoring computer in the central control room.
result under each CL. The average energy saving efficiency was about This program calculates the optimal operating point every hour, and
2.19 %. We noticed that the power load is heavy from April to October sends the setpoint to the programmable logic controller (PLC) to realize
when the climate in Taiwan becomes hotter. The cooling load of the the automatic regulation of the chiller system operating parameters.
factory is relatively high during this period. In this period, the energy This study uses the July 24, 2020 data (288 samples) that started
saving effect is quite significant. However, as mentioned previously, the running OCLR all day to compare with the July 21, 2020 data (288
obtained energy saving result is actually impractical due to that we samples) that still ran ECHWT all day. Figs. 15 and 16 depict the dis­
cannot set Tchws,i in a feasible range to accommodate the desired PLRi . tribution of operating parameters and PLR, respectively. As can be seen
from Fig. 16, there are obvious differences in the PLR operation curves of
the four chillers under OCLR control.
After we averaged the data in Fig. 15, the results are shown in

10
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

Fig. 19. Methodology of energy saving evaluation.

Table 4. We used power consumption per refrigeration (kW/RT) as an sensor data of the whole chiller system. If measurement uncertainty
efficiency index. Lower the value of (kW/RT), higher the “Efficiency” affects the power consumption of the chillers (point B in the diagram), it
and better the power saving. When we analyze Table 4, it is evident that affects the average power consumption (point A) almost equivalently.
OCLR and ECHWT provide 1 RT by consuming about 0.44722 kW and Hence, the obtained energy saving result 0.59 % is highly reliable, since
0.45168 kW, respectively. From the perspective of the efficiency index, the effect of many measurement uncertainties has been balanced out. In
the overall efficiency of OCLR is better than that of ECHWT, and the other words, even though the power consumption values may have
efficiency is improved by about 0.99 %. Since the experimental results slight error, the accuracy of power saving still remains very high. This
are averaged from 288 data, the results are sufficiently reliable. field experiment result means that the energy saving benefit of OCLR is
With the help of big data, we now present a more convincing way of credible and effective.
evaluating energy-saving efficiency. As usual, chillers using ECHWT OCLR can provide the optimal chilled water supply temperature in
runs until 13:30. Then, ECHWT is switched to OCLR. At that moment, real-time according to the weather status and the cooling load to reduce
the load distribution of the four chillers changed significantly as shown the power consumption of the chiller system. Indeed, the proposed
in the Fig. 17. approach has been implemented in several panel manufacturing fac­
We measured the power consumption of the chiller system after tories based in Taiwan running multiple-chiller system for minimizing
actually executing OCLR (14:00 ~ 19:35). The sampling rate is one data power consumption while maintaining the cleanroom environment
per five minutes. The baseline power consumption (ECHWT) is derived required for panel manufacturing.
from the historical dataset (based on 2018 data), which is established
based on the sensor data of the whole chiller system. We compared the 6. Conclusion
current measured data with the database to pick out the operating points
with similar cooling loads, dry-bulb temperature and wet-bulb temper­ This paper proposes a backward modeling approach (BMA) to ach­
ature. The power consumption values corresponding to these similar ieve practically feasible loading optimization for multiple chillers. This
parameters (may be dozens of transactions) are then averaged as the new approach solves the gap between traditional OCL methods and
baseline power consumption. Our screening conditions are that cooling practical realization of these methods. We were able to achieve signifi­
load is ±1 RT, the dry bulb is ±1 ◦ C, and the wet bulb is ±1 ◦ C. Since cant energy saving using the developed OCL regulator (OCLR). Experi­
these conditions are very strict, especially the important parameter ments have proved that even if feature data of chillers in some working
cooling load, those time points that cannot find the historical data ranges is not available, the MAPE of the OCLR using the cGAN method is
meeting the conditions have been eliminated from the comparison chart. within 1.6 % on the estimation of consumption power. In addition, the
Power consumption curves for ECHWT and OCLR, are shown in Fig. 18. power consumption model established in this study is better than to the
The average power consumption using OCLR is 1796.6 kW, whereas the traditional power consumption model established by linear regression
average power consumption of ECHWT is 1807.2 kW. This indicates methods. Based on the experimental results attained from this work, it is
that, compared to the historical data of ECHWT, OCLR has power saving estimated that OCLR can help the factory save about 81.9 MWh to 198
efficiency with the 10.6 kW (0.59 %). MWh energy per year. It is worth noting that energy saving benefits may
We summarize the evaluation methodology described above in not be limited to this. If multiple chillers in a certain application field
Fig. 19. The average power consumption (point A in the diagram) is have much variation in their working characteristic curves and the ca­
derived from the historical dataset, which is established based on the pacity, or the wider working range of Tchws,i is allowable, the energy

11
K.-Y. Lian et al. Applied Energy 327 (2022) 120102

saving efficiency of OCLR will be higher. [7] Yu J, Liu Q, Zhao A, Qian X, Zhang R. Optimal chiller loading in HVAC System
Using a Novel Algorithm Based on the distributed framework. Journal of Building
Engineering 2020;28:101044.
CRediT authorship contribution statement [8] Gao Z, Yu J, Zhao A, Hu Q, Yang S. Optimal chiller loading by improved parallel
particle swarm optimization algorithm for reducing energy consumption. Int J
Kuang-Yow Lian: Conceptualization, Formal analysis, Methodol­ Refrig 2022;136:61–70.
[9] Beghi A, Cecchinato L, Rampazzo M. A multi-phase genetic algorithm for the
ogy, Validation, Writing – review & editing, Supervision, Project efficient management of multi-chiller systems. Energy Convers Manage 2011;52:
administration, Funding acquisition, Resources. Yong-Jie Hong: 1650–61.
Conceptualization, Methodology, Investigation, Writing – original draft, [10] Chang Y-C, Chan T-S, Lee W-S. Economic dispatch of chiller plant by gradient
method for saving energy. Appl Energy 2010;87:1096–101.
Data curation, Software. Che-Wei Chang: Investigation, Data curation, [11] Ardakani AJ, Ardakani FF, Hosseinian S. A novel approach for optimal chiller
Writing – review & editing, Software. Yu-Wei Su: Investigation, Data loading using particle swarm optimization. Energy Buildings 2008;40:2177–87.
curation, Software. [12] Beghi A, Cecchinato L, Cosi G, Rampazzo M. A PSO-based algorithm for optimal
multiple chiller systems operation. Appl Therm Eng 2012;32:31–40.
[13] Huang Z, Chen J, Fu Q, Wu H, Lu Y, Gao Z, et al. HVAC optimal control with the
Declaration of Competing Interest multistep-actor critic algorithm in large action spaces. Mathematical Problems in
Engineering 2020;2020:1–12.
[14] Ning M, Zaheeruddin M. Neural network model-based adaptive control of a VAV-
The authors declare that they have no known competing financial HVAC&R system. International Journal of Air-Conditioning and Refrigeration
interests or personal relationships that could have appeared to influence 2019;27:1950006.
the work reported in this paper. [15] Gao G, Li J, Wen Y. DeepComfort: Energy-Efficient Thermal Comfort Control in
Buildings via Reinforcement Learning. IEEE Internet Things J 2020;7:8472–84.
[16] Zhou Y, Wang J, Liu Y, Yan R, Ma Y. Incorporating deep learning of load
Data availability predictions to enhance the optimal active energy management of combined
cooling, heating and power system. Energy 2021;233:121134.
The data that has been used is confidential. [17] Zhu X, Zhang S, Jin X, Du Z. Deep learning based reference model for operational
risk evaluation of screw chillers for energy efficiency. Energy 2020;213:118833.
[18] Lee D, Lai C-W, Liao K-K, Chang J-W. Artificial intelligence assisted false alarm
Acknowledgement detection and diagnosis system development for reducing maintenance cost of
chillers at the data centre. Journal of Building Engineering 2021;36:102110.
[19] Ho WT, Yu FW. Predicting chiller system performance using ARIMA-regression
This work was supported by the Ministry of Science and Technology models. Journal of Building Engineering 2021;33:101871.
(MOST), Taiwan, under Grant No. 108-3116-F-027-002-CC2- and 110- [20] Chien C-F, Chen Y-J, Han Y-T, Wu Y-C. Industry 3.5 for optimizing chiller
3116-F-027-001-. We would like to express our sincere gratitude to configuration for energy saving and an empirical study for semiconductor
manufacturing. Resour Conserv Recycl 2021;168:105247.
AU Optronics Corporation, Taiwan for giving us the opportunity to [21] Chen C-L, Chang Y-C, Chan T-S. Applying smart models for energy saving in
conduct experiments as part of our work in their company. Also, we are optimal chiller loading. Energy Build 2014;68:364–71.
grateful to their Engineers’ for introducing the workflow and machin­ [22] dos Santos Coelho L, Klein CE, Sabat SL, Mariani VC. Optimal chiller loading for
energy conservation using a new differential cuckoo search approach. Energy
eries of the company. 2014;75:237–43.
[23] Sohrabi F, Nazari-Heris M, Mohammadi-Ivatloo B, Asadi S. Optimal chiller loading
Appendix A. Supplementary material for saving energy by exchange market algorithm. Energy Build 2018;169:245–53.
[24] Ledig C, Theis L, Huszár F, Caballero J, Cunningham A, Acosta A, et al. Photo-
realistic single image super-resolution using a generative adversarial network. In:
Supplementary data to this article can be found online at https://doi. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition;
org/10.1016/j.apenergy.2022.120102. 2017. p. 4681–90.
[25] Sze V, Chen Y-H, Yang T-J, Emer JS. Efficient processing of deep neural networks:
A tutorial and survey. Proc IEEE 2017;105:2295–329.
References [26] Ioffe S, Szegedy C. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. Proceedings of the 32nd International Conference
[1] Zheng Z-X, Li J-Q, Duan P-Y. Optimal chiller loading by improved artificial fish on Machine Learning. 2015. pp. 448-56.
swarm algorithm for energy saving. Mathematics Computers in Simulation 2019; [27] Zheng Z-X, Li J-Q. Optimal chiller loading by improved invasive weed optimization
155:227–43. algorithm for reducing energy consumption. Energy Build 2018;161:80–8.
[2] Wang W, Zhou Q, Pan C, Cao F. Energy-efficient operation of a complete Chiller-air [28] Xu B, Wang N, Chen T, Li M. Empirical evaluation of rectified activations in
handing unit system via model predictive control. Appl Therm Eng 2022;201: convolutional network. arXiv preprint arXiv:00853. 2015.
117809. [29] Chang Y-C, Chen W-H, Lee C-Y, Huang C-N. Simulated annealing based optimal
[3] Trautman N, Razban A, Chen J. Overall chilled water system energy consumption chiller loading for saving energy. Energy Convers Manage 2006;47:2044–58.
modeling and optimization. Appl Energy 2021;299:117166. [30] Pascual S, Bonafonte A, Serra J. SEGAN: Speech enhancement generative
[4] Wang L, Lee EWM, Yuen RKK, Feng W. Cooling load forecasting-based predictive adversarial network. arXiv preprint arXiv:09452. 2017.
optimisation for chiller plants. Energy Build 2019;198:261–74. [31] Mirza M, Osindero S. Conditional generative adversarial nets. arXiv preprint arXiv:
[5] Catrini P, Piacentino A, Cardona F, Ciulla G. Exergoeconomic analysis as support in 14111784. 2014.
decision-making for the design and operation of multiple chiller systems in air [32] Souly N, Spampinato C, Shah M. Semi supervised semantic segmentation using
conditioning applications. Energy Convers Manage 2020;220:113051. generative adversarial network. In: Proceedings of the IEEE International
[6] Wang Y, Jin X, Shi W, Wang J. Online chiller loading strategy based on the near- Conference on Computer Vision; 2017. p. 5688–96.
optimal performance map for energy conservation. Appl Energy 2019;238:
1444–51.

12