Applied Energy 280 (2020) 115983

Contents lists available at ScienceDirect

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

A risk-based robust optimal chiller sequencing control strategy for

energy-efficient operation considering measurement uncertainties
Chaoqun Zhuang , Shengwei Wang *, Kui Shan
Department of Building Services Engineering, The Hong Kong Polytechnic University, Kowloon, Hong Kong

H I G H L I G H T S

• Risk-based robust chiller sequencing control strategy for energy-efficient operation.

• Stochastic decision-making scheme is developed to evaluate risks of chillers in operation.
• Probabilistic simplified physical cooling load and capacity models are developed.
• Probabilistic models are calibrated using Bayesian Markov chain Monte Carlo method.
• Proposed strategy could improve chiller energy efficiency and ensure thermal comfort.

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

Keywords: Proper and reliable control of central chilling systems with multiple chillers is crucial to save energy and enhance
Risk-based control energy efficiency. The conventional total-cooling-load-based chiller sequencing control strategies determine
Measurement uncertainty switching (on/off) thresholds according to building instantaneous cooling load and chiller maximum cooling
Chiller sequencing
capacity. However, due to the existence of measurement uncertainties and ever-changing operating conditions,
Bayesian calibration
Robust optimal control
optimal switching points often deviate significantly from predefined thresholds. To deal with these challenges
and uncertainties, a risk-based robust optimal chiller sequencing control strategy is proposed to improve the
robustness and energy efficiency of chillers in operation. As the core of the control strategy, an online stochastic
decision-making scheme, which is developed to optimize chiller staging based on quantified risks. The risk of
failure to achieve expected operation performance by switching on/off a chiller is evaluated through analyzing
the probabilistic fused cooling load and the probabilistic chiller maximum cooling capacity, based on Bayesian
calibration of cooling load and capacity models. The best switching points can therefore be identified in a sto­
chastic approach. The results of case studies show that the proposed strategy can improve the reliability and
robustness of chiller sequence operation. Compared with the conventional strategy, the switching frequency was
decreased by more than 54%, and the energy use of central cooling systems can be reduced by 2.8% without
sacrificing thermal comfort.

control is a control function that determines the number of chillers to be

1. Introduction staged on/off in a given load condition [2]. Optimal chiller sequencing
control should provide buildings with sufficient cooling while avoiding
1.1. Background energy wastage. Excessive operation of chillers could certainly fulfill
demanded cooling load, but would consume more electricity. However,
Central cooling systems with multiple chillers are widely used in inadequate operation of chillers would fail to provide enough cooling
modern commercial buildings to provide cooling for indoor spaces. As a and therefore cause thermal discomfort.
major energy consumer, chillers can consume over half of the total Chiller sequencing control strategies can be categorized into direct or
building electricity use [1], leading to considerable interest in the po­ indirect methods according to indicators of building instantaneous
tential for energy savings. Proper control of multiple chillers is crucial to cooling loads [3]. The direct method determines the chiller stages based
save energy and enhance system energy efficiency. Chiller sequencing on cooling load directly by measuring supply and return water

* Corresponding author.
E-mail addresses: [email protected] (C. Zhuang), [email protected] (S. Wang), [email protected] (K. Shan).

https://doi.org/10.1016/j.apenergy.2020.115983
Received 24 June 2020; Received in revised form 15 September 2020; Accepted 4 October 2020
Available online 19 October 2020
0306-2619/© 2020 Elsevier Ltd. All rights reserved.
C. Zhuang et al. Applied Energy 280 (2020) 115983

Nomenclature SNtot total switching number of chillers

T temperature (◦ C)
AHU air-handling unit TH threshold (kW)
CPg gaseous refrigerant specific heat at constant pressure, kJ/ vw water volumetric flow rate (L/s)
(K⋅kg) Z compressibility factor of the refrigerant
CPl liquid refrigerant specific heat at constant pressure, kJ/
(K⋅kg) Greek letters
Crated chiller rated capacity (kW) Ф cumulative density function
cw specific thermal capacity of water (kJ/kg⋅◦ C) δ constant part of the electromechanical losses (kW)
CV(RMSE) coefficient of variation of the root mean square error α loss factor of the variable part of electromechanical losses
d dead band ν maximum volumetric flow rate of the gaseous refrigerant
Ecom electrical power of chillers (kW) (L/s)
E energy consumption (kWh) ε aggregated systematic (bias) error (kW)
hfg latent heat of vaporization at the reference state pressure θ unknown parameter
(kJ/kg) η normalization constant
j, k number of samples β risk threshold
M measured value ζ total integrated time (h)
MCMC Markov Chain Monte Carlo Subscripts
N number of chillers in operation chi chiller
Nmean chiller mean stages ct cooling tower
Pcd condensing pressure (kPa) dm direct measurement
Pev evaporating pressure (kPa) im indirect measurement
qfu fused cooling load (kW) pum pump
qmax maximum cooling capacity (kW) rtn return chilled water
Qfu probabilistic fused cooling load (kW) sup supply chilled water
Qmax probabilistic maximum cooling capacity (kW) tot total
R gas constant
Rf risk/failure probability

temperature as well as total chilled water flow. The indirect methods Conventionally, the rated cooling capacity provided by manufacturers is
determine the chiller stages based on indirect indicators of cooling load, used as a switching threshold while the difference between the rated and
such as the chilled return water temperature, bypass water flow, chiller real cooling capacity of chillers can certainly affect chiller operating
vane opening and chiller current/power. The total-cooling-load-based efficiency. To consider the effects of variations in chiller cooling ca­
chiller sequencing control strategies (i.e. direct methods) are consid­ pacity, Liu et al. [12] used iteration loops to estimate the maximum
ered the most promising strategy in principle as the indirect indicators cooling capacity of individual chillers based on the energy balance be­
may not be proportional to the cooling load [4]. It is straightforward and tween chillers and cooling towers. In addition, data-driven models were
convenient to measure cooling loads by the product of the water flow adopted by Shan et al. [13] to predict maximum cooling capacity under
rate and differential chilled water temperature. However, site studies varying chiller operating conditions. Due to the propagation and inter­
showed that the total-cooling-load-based chiller sequencing control action of different uncertainty sources in operation, the evaluation of the
strategy does not always ensure the proper operation of systems because aggregated uncertainty of a multiple-chiller plant becomes complicated.
the differential temperature cannot be measured accurately [5]. The Uncertainty information is critical in supporting the decisions for
accuracy of the direct measurement of cooling loads may be significantly achieving optimal design/control objectives [14]. Decision making
influenced by measurement errors, making no sense for incorporating an based on stochastic programming is one of the effective approaches to
energy-efficient chiller sequencing control [6]. For instance, an error of improve the estimation accuracy and decision quality in presence of
0.4 ◦ C in the measurement of either supply or return chilled water uncertainty [15]. The main advantage of stochastic programming over
temperature can lead to an error rate of up to 10% in total cooling load its deterministic counterpart is that the optimal objective in decision-
measurement, if the differential temperature is 4 ◦ C. A survey of 30 making process can be obtained in an expected value sense [16]. In
central chiller plants conducted in Hong Kong also indicated that the addition, this method can also introduce flexibility in decision making
measurements of water temperature and flow rate were not reliable/ according to decision-maker attitudes (e.g. risk-neutral or risk-averse) in
accurate enough for proper sequence operation [7]. responding to possible changes that are uncertain or as yet unknown
Various chiller sequencing control strategies have been developed to [17]. With respect to addressing uncertainties, risk assessment with
handle these uncertainties in cooling load measurement. Sun et al. [8] advanced stochastic models [18] has been widely used for balancing
utilized fused measurements of building cooling loads to improve the between acceptable levels of risk and the costs of further risk reduction
reliability of chiller sequencing control, by combining the complemen­ [19]. For instance, a multi-stage stochastic optimization model, which
tary advantages of both direct and indirect cooling load measurement. involves risk management considering the variability of the uncertain
Huang et al. [9] developed a strategy of fusing available redundant parameters, was developed by Cano et al. [20] for optimal planning of
measurements to improve the accuracy of cooling load measurements building energy systems. Chou and Ongkowijoyo [21] proposed a risk-
and reduce measurement uncertainties using a model-free method. Liao based stochastic graphical matrix model to determine renewable en­
et al. [10] proposed a practical way to enhance the robustness of chiller ergy system schemes with improved reliability. Mavromatidis et al. [22]
sequencing control by making use of the complementarity of different proposed a two-stage risk-based stochastic program for sizing cost-
load indicators. Apart from the uncertainties in cooling load measure­ optimal distributed energy systems considering the uncertain heat and
ments, the maximum cooling capacity of individual chillers is also electricity demands, solar radiation and electricity price. For existing
difficult to determine as it varies with operating conditions [11]. studies, stochastic decision-making approaches are generally developed

2
C. Zhuang et al. Applied Energy 280 (2020) 115983

for the applications in the design or plan stage, while there is lack of (1) The chiller sequencing control strategy based on risk assessment
feasible online decision-making approach to ensure robust chiller with a stochastic manner, to the authors’ best knowledge, is
sequencing control considering measurement uncertainties. proposed for the first time to address the measurement un­
certainties, improving the reliability and robustness of chillers in
2. Novelty and main contributions operation.
(2) A data fusion approach, as a key element in the online stochastic
To reduce/consider the impacts of measurement uncertainties, pre­ decision-making scheme, is proposed to merge the two probabi­
vious studies mainly focused on improving the accuracy of cooling load listic cooling load measurements, so as to remove the outliers and
measurement and chiller cooling capacity estimations. The operating reduce the impact of the measurement uncertainties on the
number of chillers is determined by comparing the deterministic values determination of chiller staging.
of cooling load and chiller maximum cooling capacity. Although a sto­ (3) A Bayesian Markov Chain Monte Carlo method is adopted to
chastic chiller sequencing control strategy is used by Li et al. [23], by quantify the unknown parameters of probabilistic models, which
quantifying the uncertainties associated with cooling load measurement can take account of the “aggerated uncertainties” that exist in the
and chiller cooling capacity with predefined normal/uniform distribu­ estimation of the cooling load/capacity.
tions, the parameter settings of the distributions would still significantly
affect final decision making. 3. Risk-based robust optimal chiller sequencing control strategy
This study therefore proposes a risk-based robust optimal chiller
sequencing control strategy for achieving the energy-efficient operation 3.1. Basic idea and major innovation
of chillers considering measurement uncertainties. The main innovation
of this strategy is the development of an online stochastic decision- Fig. 1 shows the schematic diagram of a typical concerned chilled
making scheme to optimize chiller switching based on the quantified water system, where its water loop is decoupled into primary and sec­
risks (i.e. failure in achieving expected operation performance). The ondary distribution loops balanced by a bypass line. In the primary loop,
risks by switching on/off a chiller are quantified and used for decision each chiller is interlocked with a constant-speed chiller water distribu­
making, considering changes in operating conditions as well as mea­ tion pump and a cooling water distribution pump. The chilled supply
surement uncertainties. Probabilistic simplified physical models are water from chillers is mixed up, flowing into terminal air handling units
developed to determine the distributions of cooling loads and maximum (AHUs) in the secondary loop to provide cooling for the spaces. An
cooling capacities for online risk quantification. The proposed control online decision-making scheme is used to determine chiller staging ac­
strategy is developed by addressing the following two tasks: (1) offline cording to building instantaneous cooling loads and chiller cooling
uncertain parameter identification in probabilistic models for cooling capacities.
load/capacity estimations; (2) an online risk-based decision-making The decision-making schemes of the conventional total-cooling-load-
scheme to determine optimal chiller stages. The uncertain parameters of based and proposed risk-based chiller sequencing control strategies are
the probabilistic models are identified using the Bayesian Markov Chain presented in Fig. 2. For the conventional strategy (Fig. 2A), the thresh­
Monte Carlo (MCMC) method. The risks of switching on/off a chiller are olds for switching on the (N + 1)th chiller and switching off the Nth
quantified by analyzing the probabilistic fused cooling load and the chiller, denoted by THN + 1 onand THN off, are normally deterministic,
probabilistic chiller maximum cooling capacity. To verify the proposed based on the requirement that the maximum cooling capacity provided
strategy, case studies are conducted in the tallest building in Hong Kong. by chillers should exceed the actual cooling load with minimum energy
The energy performance and control robustness of the proposed strategy cost. A dead band d is considered to prevent frequent switching in case
are investigated and compared with the conventional total-cooling-load the cooling load varies within a small range near the switching points.
based strategy. The cooling load thresholds for switching on/off a chiller are defined by
The original contributions of this study are briefly summarized as Eq. (1) when the chillers are identical. Here, N is the number of chillers
follows: in operation. Crated is the rated capacity of chillers (kW). The optimal
number of operating chillers can then be determined by Eq. (2). The

Fig. 1. Schematic diagram of a typical chilled water system.

3
C. Zhuang et al. Applied Energy 280 (2020) 115983

Fig. 2. The basic idea of the proposed risk-based chiller sequencing control and comparison with the conventional strategy.

main disadvantage of the conventional strategy is that, under off-design

conditions, its switching thresholds may deviate far from the best
switching points due to measurement uncertainties and changes in
operating conditions (varying chiller maximum cooling capacities).
{ N
TH off = (N − 1)⋅[Crated ⋅(1 − d)]
(1)
TH N+1
on = N⋅[Crated ⋅(1 + d) ]

{
N − 1, Q < TH Noff
Nnext,con = (2)
N + 1, Q > TH N+1
on

The development of the online stochastic decision-making scheme to

determine the chiller staging is the major innovation of the proposed
risk-based chiller sequencing control strategy. For the proposed strategy Fig. 3. Probabilistic cooling load and maximum cooling capacity for risk
(Fig. 2B), the switching points are determined based on quantifying both quantification.
the risk (i.e. failure in achieving the expected operation performance)
when switching off Nth chiller (Rf, N− 1) and the risk when operating Monte Carlo (MCMC) method [24] is used to calibrate the probabilistic
under current N chillers (Rf, N). The optimal operating number of chillers models using in-situ measurements. The second task is to quantify the
can then be determined through Eq. (3), where βon and βoff (β ∈ [0, 1]) risks of switching on/off a chiller for online decision making. A data
are user-defined thresholds and Ф is the cumulative density function fusion approach is developed to fuse the probabilistic direct and indirect
(CDF). The risk thresholds (βon and βoff) should be properly selected cooling loads for outlier removal. The risks are quantified by analyzing
according to the requirements of building thermal comfort and energy the probabilistic fused cooling load and probabilistic chiller maximum
efficiency. For a comfort-oriented system requiring sufficient cooling, cooling capacity. The details to address each task are illustrated in
βon and βoff should be set to relatively small values. For an energy- Sections 2.3 and 2.4, respectively.
oriented system requiring high energy efficiency, βon and βoff should
be set to larger values. By adopting the proposed decision-making
scheme, the best switching points can always be identified under mea­ 3.3. Calibration of probabilistic cooling load and maximum cooling
surement uncertainties and different operating conditions. Here, Rf, N− 1 capacity models
and Rf, N are evaluated online by analyzing the probabilistic fused
cooling load (Qfu=[qfu,1, qfu,2, …, qfu,k]) and the probabilistic maximum The building cooling loads can be computed through direct and in­
cooling capacity of N-1 and N chillers (Qn-1 max = [qn-1 max,1, qn-1 direct measurements. Each method has benefits [8]. The direct mea­
max,2, …, qn-1 max,k] and Qn max = [qn max,1, qn max,2, …, qn max,k]) surement of building cooling load (qdm) can be formulated as shown in
as presented in Fig. 3. The quantification of fused cooling loads and Eq. (4). cw is the specific thermal capacity of water (kJ/kg⋅◦ C). vw is the
chiller maximum capacities is further elaborated in Section 2.4. water volumetric flow rate (L/s). Trtn and Tsup are chilled return and
{ ( ) supply water temperature (◦ C) respectively. However, qdm is highly
N − 1, Rf ,N− 1 = Φ Qfu > QN− susceptible to measurement uncertainties (especially those associated
1
≤ βoff
(3)
max
Nnext,new = ( )
N + 1, Rf ,N = Φ Qfu > QNmax ≥ βon with Trtn and Tsup). To improve the quality of building cooling load
measurements, an indirect method using a probabilistic indirect cooling
load model (i.e. model-based cooling load measurements), relating
3.2. Outline of risk-based chiller sequencing control strategy building cooling load to measurements from the chiller side, can be
beneficial in improving the reliability of cooling load measurements.
The outline of the proposed risk-based chiller sequencing control )
qdm = cw ⋅ρw ⋅vw ⋅(Trtn − Tsup (4)
strategy is shown in Fig. 4, which consists of two major tasks to be
addressed. The first task is to identify the uncertain parameters of Additionally, the maximum cooling capacity of a chiller often de­
simplified physical models for estimating the cooling load and chiller viates from its rated cooling capacity, affecting the decision making of
maximum cooling capacity in an indirect way. A Bayesian Markov Chain chiller sequence operation. To consider varying chiller cooling capacity

4
C. Zhuang et al. Applied Energy 280 (2020) 115983

Fig. 4. Framework of the proposed risk-based chiller sequencing control strategy.

under different operating conditions, a probabilistic maximum cooling systematic (bias) error term (kW), which can be determined under
capacity model, using the full-load operation data from the chiller side, different numbers of operating primary chilled water pumps. R is the gas
is used to calculate the maximum chiller cooling capacity without an constant. Z is the compressibility factor of the refrigerant. α, δ, v and ε
iteration loop. are unknown parameters to be determined using in-situ measurements.
In this study, two probabilistic simplified physical models, an indi­ It is worth noting that the unknown parameters (e.g. α, δ and v) of the
rect cooling load model and a maximum cooling capacity model, are simplified physical models used in [8] are deterministic values while the
developed based on the two deterministic models in reference [8]. The parameters used in this study were quantified with certain distributions.
measurement uncertainties are aggregated into the probabilistic models Compared with direct cooling load measurements, the indirect
by identifying the model parameters as elaborated below. measurements of cooling loads qim are generally reliable. As the
condensing temperature and evaporating differential temperature (Ṫev-
3.3.1. Probabilistic simplified physical models for cooling load/chiller Ṫ cd) have a much larger scale (i.e. over 20 ◦ C) than return and supply
cooling capacity estimation water differential temperature (Ṫ rtn-Ṫ sup), the effects of associated
The indirect measurements of cooling load qim and maximum cooling measurement noises on indirect measurements (Eq. (5)) are much
capacity with N operating chillers (qn max) can be computed using two smaller than that on the direct measurements (Eq. (4)).
probabilistic simplified physical models as shown in Eqs. (5)–(6).
∑
N
( ) 3.3.2. Identification of model parameters using Bayesian Markov Chain
qim = fi Ecom,i , Tcd,i , Tev,i Monte Carlo method
i=1
[ ] The Bayesian MCMC method is employed to calibrate uncertain
∑ parameters in the model using the in-situ data, containing two stages:
N
Ecom,i − δi ( )
= ( ) CPl ⋅Tcd,i − hfg − CPg ⋅Tev,i + εN (5)
i=1 αi ⋅CPg Tev,i − Tcd,i Bayesian inference and MCMC sampling. The Bayesian inference derives
the posterior distributions for the calibrated parameters (i.e. α, δ, ν and
∑
N
( ) ε) using prior information and observed data, while MCMC sampling is
qNmax = fi Pev,i , Tcd,i , Tev,i used to generate equivalent samples from the posterior distributions.
The equivalent samples represent the numerical approximations of the
i=1
[ )]
∑ posterior distributions, and thus they can be used to represent the fea­
n
ν ⋅P
= ( i ev,i ) (hfg + CPg ⋅Tev,i − CPl ⋅Tcd,i + εN (6)
i=1 R⋅Z⋅ 273.15 + Tev,i tures of unknown parameters.
Bayesian inference: Given a model y = f (θ) and the measured data of
Here, α is the loss factor of the variable part of electromechanical y, where θ is the unknown parameter to be determined and y is the
losses. δ is the constant part of the electromechanical losses (kW). hfg is output, the Bayesian inference deduces the posterior probability P (θ|y)
the latent heat of vaporization at the reference state pressure (kJ/kg). through Eq. (7). Here P (y|θ) is the likelihood of the measured data y
CPg is the gaseous refrigerant specific heat at constant pressure in kJ/ occurring given the input parameter θ. P (θ) is the prior probability that
(K⋅kg). CPl is the liquid refrigerant specific heat at constant pressure in describes the probability of θ occurring computed before the collection
kJ/(K⋅kg). Ecom is the electrical power of chillers (kW), which can be of in-situ data. P (y) is a normalization constant. In this study, the
measured directly. Tcd and Tev are condensing temperature and evapo­ cooling load qim and maximum cooling capacity of chillers qmax are the
rating temperature derived according to the corresponding condensing output y. qim and qmax are calculated through Eq. (4) using the in-situ
pressure Pcd and evaporating pressure Pev (kPa). ν is the maximum data of vw, Trtn and Tsup. The in-situ data of Ecom, Pev, Tcd and Tev as
volumetric flow rate of the gaseous refrigerant (L/s). ε is an aggregated

5
C. Zhuang et al. Applied Energy 280 (2020) 115983

well as the corresponding qim and qmax values are used for generating the probability density function of Qn max. The probability density function
likelihood based on Eqs. (5)–(6). Notably, the chiller full-load operation of X2 (fX2(X2)) is the convolution of ffu and f n max as described by Eq.
data (i.e. chiller vane opening/current higher than high thresholds) are (11), where X2 = Qfu-Qn-1 max and fn-1 max is the probability density
selected for calculating qmax, while both the chiller part-load and full- function of Qn-1 max.
load operation data are selected for calculating qim. ∫
fX1 (X1 ) = ffu (X1 + y)⋅(y)dy (10)
P(θ|y) = P(y|θ)⋅P(θ)/P(y)∝P(y|θ)⋅P(θ) (7)
Markov chain Monte Carlo (MCMC) sampling: MCMC, specifically the ∫
Metropolis-Hastings algorithm [25], is employed to draw random sam­ fX2 (X2 ) = n− 1
ffu (X2 + y)⋅fmax (y)dy (11)
ples from the joint multivariate posterior distribution. As a result, a
sequence of samples with size m, {θ (1), θ (2), …, θ (m)} which approxi­ 4. Building air-conditioning system and arrangement of online
mates the theoretical posterior probability density functions (pdfs) of validation tests
calibration parameters can then be collected. More details for the MCMC
sampling method are elaborated in the literature [26]. 4.1. Description of the air-conditioning system
It is worth noting that, since two probabilistic models were trained
using the same sources of measurements, the risks of switching on/off a The central cooling system concerned in this study is a complex
chiller (Eq. (3)), determined based on the differential of cooling loads primary-secondary chilled water system in the tallest building in Hong
and cooling capacities, can always be reliably quantified in the face of Kong, where the cooling is dominated throughout a year [27]. The
measurement uncertainties. building has a height of 490 m and consists of 108 floors with around
321,000 m2 floor area. Table 1 shows the specifications of the main
3.4. Online stochastic decision-making scheme for chiller sequencing equipment in the central cooling system. Six centrifugal chillers with a
control rated coefficient of performance (COP) of 5.17 are equipped, providing
the designed chilled water supply temperature of 5.5 ◦ C. Each chiller is
The online stochastic decision-making scheme for determining the interlocked with a constant-speed chilled water pump (rated flow rate of
chiller stages involves three steps as follows. 375 L/s) and a constant-speed cooling water pump (rated flow rate of
The first step is to obtain the samples of indirect and direct cooling 410 L/s). The heat rejection system consists of eleven cooling towers
load measurements as well as the maximum cooling capacities of with a total design capacity of 51,709 kW.
chillers. At each control interval, the indirect cooling loads (Qim = [qim,1,
qim,2, …, qim,k]), the maximum cooling capacities with current operating
4.2. TRNSYS-MATLAB co-simulation testbed and test conditions
chillers (Qnmax = [qnmax,1, qnmax,2, …, qnmax,k]) and the maximum cooling
capacities when switching off a chiller (Qn-1max = [qn-1max,1, qn-1max,2, …,
To test the effectiveness of the proposed strategy, a virtual simulation
qn-1max,k]) are calculated/sampled k times adopting the probabilistic
platform was constructed based on the configuration of the studied
simplified physical models (Eqs. (5)–(6)). j samples of direct cooling
central cooling system using a Transient Simulation Program TRNSYS
load measurements (Qdm = [qdm,1, qdm,2, …, qdm,j]) within the control
18 [28]. To take advantage of its powerful computational capabilities,
interval were also selected, assuming the building cooling load is rela­
MATLAB was used to program the supervisory controller (i.e. risk-based
tively stable in a short period.
chiller sequencing control strategy) and was embedded in TRNSYS. The
In the second step, to enhance the robustness of building cooling load
dynamic processes of hydraulics, heat transfer, water flow/pressure
measurements, a data fusion approach is used to capitalize on the ben­
balancing, energy conservation and control were simulated for the
efits of direct and indirect cooling load measurements, removing the
entire system. The models used in the test platform were calibrated using
measurement errors and model errors simultaneously. The probability
on-site operational data [29,30]. A simplified global air handling unit
density function of fused cooling loads (ffu) is computed by the product
(AHU) was built to provide cooling for a thermal building. The “multi-
of two distributions (i.e. direct and indirect cooling load measurements)
zone building model” type 56 (i.e. the thermal building) in TRNSYS was
as shown in Eq. (8), where fdm and fim are probability density functions of
employed to simulate the thermal behavior of the selected building. The
Qdm and Qim respectively and η is a normalization constant (expressed by
building model in type 56 is a non-geometrical balance model with one
Eq. (9)). The fused cooling loads (Qfu=[qfu,1, qfu,2, …, qfu,k]) are then
air node per zone, representing the thermal capacity of the zone air
sampled k times given ffu. Notably, if the direct measurements deviate
volume and capacities which are closely connected with the air node
far from the indirect measurements (i.e. fdm⋅fim is a null distribution),
[28]. The envelope construction and thermal properties were set satis­
indicating the indirect measurements may suffer significant model er­
fying the requirements of the building design description and local en­
rors (usually occurring at the very start of chiller switching on/off), the
ergy efficiency standard [31]. The building’s thermal performance and
ffu is set as fdm. The use of fused cooling loads instead of direct/indirect
dynamics under the influences of weather, occupancy, and air-
cooling loads has two benefits. (1) When a direct measurement suffers
conditioning systems were then characterized.
incidentally from large random errors (marked as “outlier”) in a control
The actual cooling load data (without measurement uncertainties)
interval, this outlier can be removed as its probability of occurrence is
near zero in fim. (2) When an indirect measurement suffers from large
model errors, this outlier can be removed as its probability of occurrence Table 1
is near zero in fdm. Specifications of main equipment in the system.
{ Equipment Number Rated Flow rate Power Head
η⋅fdm (x)⋅fim (x)if : fdm (x)⋅fim (x) ∕
= null capacity (kW) (L/s) (kW) (kPa)
ffu (x) = (8)
fdm (x)if : fdm (x)⋅fim (x) = null
Chiller 6 7,230 – 1,346 –
Cooling tower 6 5,234 – 152 –
1 A
η=∫ (9)
fdm (x)⋅fim (x)dx Cooling tower 5 4,061 – 120 –
B
In the third step, the chiller stages can be determined based on Chilled water 6 – 345.0 126 31.6
quantified risks through Eq. (3), using fX1(X1) and fX2(X2). Here, the pump
probability density function of X1 (fX1(X1)) is the convolution of ffu and f Cooling water 6 – 410.1 202 41.6
pump
n max as expressed by Eq. (10), where X1 = Qfu-Qn max and f n max is the

6
C. Zhuang et al. Applied Energy 280 (2020) 115983

used in the case study were generated using a typical cooling load profile Table 2
of the selected building in 7 days, as presented in Fig. 5. To test the Sensor noises and biases introduced for case studies.
robustness of the proposed control strategy, measurement uncertainties Measurement Unit Bias Noise
(including biases and noises) were set for different types of sensors in
Supply water temperature ◦
C Case 1: ¡0.4; Case 2: 0.4 N (0, 0.1)
two cases, as summarized in Table 2. Case 1 represents the conditions Return water temperature ◦
C Case 1: 0.2; Case 2: ¡0.2 N (0, 0.1)
where the measured cooling loads are higher than actual cooling loads Supply water flowrate L/s Both cases: 20 N (0, 8)
while Case 2 represents the conditions where the measured cooling Chiller condensing temperature ◦
C Both cases: − 0.3 N (0, 0.1)
loads are lower than actual cooling loads. Chiller evaporating temperature ◦
C Both cases: 0.2 N (0, 0.1)

5. In-situ validation of calibrated probabilistic cooling load and

capacity models Table 3
In-situ data type and variables.
In this study, 525,600 sets of in-situ measurements (listed in Table 3), No. Measurement Unit Symbol
obtained from the central cooling systems of the selected high-rise 1 Return chilled-water temperature ◦
C Trtn
building with the 1-minute interval in 2019, were used to calibrate 2 Supply chilled-water temperature ◦
C Tsup
and validate the models. 3 Total water volumetric flowrate L/s vw
The data of Groups 1–21 from 1 July to 31 July were used to identify 4–9 Electrical power of chillers 1–6 kW Ecom,i (i = 1,2,3,4,5,6)
10–15 Condensing pressure of chillers 1–6 C Pcd,i (i = 1,2,3,4,5,6)
the parameters of probabilistic indirect cooling load models, while these
◦

16–21 Evaporating pressure of chillers 1–6 ◦

C Pev,i (i = 1,2,3,4,5,6)
groups of data from 3 Aug to 10 Aug were used to validate these models. 22–27 Current of chillers 1–6 Ampere Ii (i = 1,2,3,4,5,6)
The chiller full-load operation data in the first half-year were used to 28–35 Vane opening of chillers 1–6 % VA,i
identify the parameters of maximum cooling capacity models while the
latter half-year data were used to validate these models. The chiller full-
load operation data were selected by identifying whether the current Table 4
and vane opening of the chillers were larger than 70 amperes and 95% Prior and posterior distributions of uncertain parameters.
respectively. These data were considered observation evidence and were
No. Parameter Unit Prior distribution Posterior distribution
used for producing the likelihood functions. The refrigerant used in the
chiller is R134a. The constants used in the model are R⋅Z: 73.41 kJ/ Distribution Mean Mean [25%, 75%]

(kg⋅K); hfg: 197.9 kJ/kg; CPg: 0.89 kJ/(kg⋅K) and CPl: 1.27 kJ/(kg⋅K). 1 α1 – U [0.31,0.41] 0.36 0.36 [0.34,0.38]
A total of 21 uncertain parameters in six indirect cooling load models 2 δ1 kW U [621,841] 731 734 [713,750]
3 L/s U [1990,2692] 2341 2361 [2322,2402]
and six chiller maximum cooling capacity models need to be identified. ν1
4 α2 – U [0.26,0.36] 0.31 0.31 [0.29,0.32]
Here, αi, δi and vi were set (i = 1,2,3,4,5,6) as uniform distributions while 5 δ2 kW U [610,825] 718 716 [699,745]
εn (n = 1,2,3) were set as normal distributions. Notably, the maximum 6 ν2 L/s U [2053,2777] 2415 2362 [2323,2401]
value of n is 3 because the maximum operating number of chillers was 3 7 α3 – U [0.44,0.60] 0.52 0.52 [0.49,0.53]
in the operating period. Due to a lack of prior knowledge of those un­ 8 δ3 kW U [485,657] 571 568 [553,582]
9 ν3 L/s U [1958,2650] 2304 2288 [2171,2407]
certain parameters, the deterministic values calculated by the nonlinear 10 α4 – U [0.33,0.44] 0.38 0.38 [0.36,0.40]
least-square method [32] based on Eqs.5–6 with in-situ measurements, 11 δ4 kW U [559,757] 658 656 [640,671]
were assigned as the mean values of the prior distributions. The lower/ 12 ν4 L/s U [1557,2106] 1832 1823 [1791,1863]
upper limits of αi, δi and vi were assumed to be a deviation of 15% from 13 α5 – U [0.30,0.40] 0.35 0.36 [0.34,0.38]
14 kW U [579,783] 681 683 [668,698]
their corresponding mean values. The standard deviation of εn was set as δ5
15 ν5 L/s U [1927,2608] 2268 2302 [2253,2249]
a 5% random error [33]. 16 α6 – U [0.43,0.59] 0.51 0.51 [0.49,0.53]
The detailed information for prior distributions of uncertain pa­ 17 δ6 kW U [453,613] 533 534 [520,546]
rameters can be found in Table 4. A total of 50,000 MCMC simulations 18 ν6 L/s U [453,613] 2315 2326 [2190,2421]
were performed while the first 10,000 samples were discarded as a burn- 19 ε1 kW N [536,27] 536 542 [528,557]
20 ε2 kW N [977,49] 977 978 [960,995]
in period. Fig. 6 shows the calibrated distributions of 21 uncertain pa­
21 ε3 kW N [1576,78] 1555 1576 [1550,1601]
rameters. Generally, the posterior distribution becomes irregular if the
prior distribution follows a uniform distribution, while the posterior Note: U represents a uniform distribution and N represents a normal
distribution.
distribution nearly follows a normal distribution if the prior distribution
follows a normal distribution. Table 4 also lists the characteristic values
of the calibrated distributions to be used for comparing with their prior
distributions. The mean values of parameters did not change much after

Fig. 5. The actual cooling loads used in the test.

7
C. Zhuang et al. Applied Energy 280 (2020) 115983

Fig. 6. Distributions of identified parameters of simplified probabilistic physical models.

calibration, but the occurrence probability of the values shifted and data. Here Mi and Ei are the measured and model estimated mean
dramatically (e.g. especially for the parameters following uniform dis­ values at ith sampling instant respectively. The measured average values
tributions). Overall, the calibration process changed the prior informa­ of the cooling load/capacity are denoted by M.
tion and made the distributions of parameters more informative. ∑n
To show the predictive ability of the probabilistic simplified physical (Mi − Ei )2
CV(RMSE) = i=1 (11)
models with calibrated parameters (posterior information), the proba­ M
bilistic indirect cooling load and maximum cooling capacity, calculated By calculating the index of agreement between model estimated
by the calibrated models (i.e. 2,000 samples at each sampling instant), mean and measured cooling load/capacity, it is found that CV(RMSE)
are compared with in-situ data as shown in Figs. 7 and 8. The figures on for probabilistic cooling/capacity models were 7.5% and 2.9%, respec­
the right present the probability distributions of the indirect cooling load tively. It can be seen that most of the measured load/capacities fell in the
and maximum cooling capacity in a selected sampling instant. The co­ 95% prediction intervals. This confirms that the calibrated probabilistic
efficient of variation of the root mean square error CV(RMSE) [34] was simplified models are satisfactory, and can be used to reliably estimate
used to evaluate the similarity between the model estimated values and the cooling load/capacity.
actual values, as shown in Eq. (11). The range of CV(RMSE) lies between
0 and 100%, with lower values signifying a good fit between the model

Fig. 7. Indirect measurements of cooling loads using a probabilistic indirect cooling load model vs direct measurements (3–10/Aug/2019).

8
C. Zhuang et al. Applied Energy 280 (2020) 115983

Fig. 8. Estimated maximum cooling capacity using maximum cooling capacity model of chiller 1 vs real measurements (July-Dec/2019).

6. Online operation performance and validation of proposed • Risk thresholds (βon and βoff) were set as 0.3 and 0 respectively for the
sequence strategy proposed strategy.

The control performance and energy efficiency of the central cooling Four performance indicators, including the total energy consumption
system using the conventional strategy and proposed risk-based chiller ECtot, total switching number SNtot, chiller mean stages Nmean and total
sequencing control strategy were evaluated in a virtual simulation integrated time (ζ0.5) where the thermal building’s average temperature
platform. The normal operational data were generated for re-calibrating deviated 0.5 ◦ C from its setpoint (23 ◦ C), were used to evaluate the
model parameters by simulating the central cooling systems over a wide performance of the chiller sequencing control strategies. The total en­
range of operating conditions considering measurement uncertainties. ergy consumption ECtot includes the energy use of chillers (ECchi),
The chiller sequencing control algorithm was used in the case study constant-speed pumps (ECpum) and cooling towers (ECct).
with the following settings:
6.1. Results of Case 1: measured cooling loads > actual cooling loads
• Both the simulation time step and the sampling interval of mea­
surements were set 1 s, while the control interval to determine the Fig. 9 shows the comparison of direct, indirect and fused measure­
chiller switching on/off was 5 min. ments of cooling loads in Case 1. There are few outliers in the direct and
• The samples (j) of direct cooling load measurements were set as 60 (i. indirect measurements, also marked in the figure. Two outliers in the
e. 1-min data before the control execution). The samples (k) of in­ direct measurements (i.e. around 83.5 h and 112 h) mainly resulted from
direct and fused cooling loads were set as 2,000. large random errors while the outliers in the indirect measurements
• The dead band d was set as 0.05 (Eq. (1)) for the conventional mainly resulted from model errors (e.g. around 2.3 h and 32.8 h),
strategy. appearing during very early periods of chiller staging. The use of fused
measurements can take advantage of direct and indirect measurements.

Fig. 9. Comparison on direct, indirect and fused measurements of cooling load.

9
C. Zhuang et al. Applied Energy 280 (2020) 115983

Fig. 10. Control performance of the proposed and conventional strategies in Case 1: (A) chiller operating number (B) indoor temperature of the thermal building.

The outliers or the strikes were efficiently removed, which can avoid As in Case 1, the number of the operating chillers varied frequently
unnecessary chiller switching. when adopting the conventional strategy, while the switching frequency
The variation in the number of operating chillers adopting the con­ was significantly reduced when adopting the proposed strategy. How­
ventional and proposed strategies are compared and presented in ever, in contrast to Case 1, the operating number of chillers adopting the
Fig. 10A. It can be seen that the number of chillers in operation varied conventional strategy was no higher than that of adopting the proposed
frequently when adopting the conventional strategy. However, the strategy in the test period, due to underestimation of cooling loads.
switching frequency was significantly reduced when adopting the pro­ Fig. 12B shows the indoor temperature under the two different strate­
posed strategy. In addition, the operating number of chillers adopting gies. The reduction of chiller operating numbers in some periods resul­
the conventional strategy was no lower than those adopting the pro­ ted in temperature rise (from 133 h to 138 h) due to insufficient cooling
posed strategy in the test period, due to overestimation of the cooling supply under the conventional strategy, while the indoor temperature
loads. Fig. 10B shows the indoor temperature of the two strategies. was well controlled under the proposed strategy. As also listed in
When adopting the proposed strategy, the reduction of chiller operating Table 6, compared with the conventional strategy, the total switch
numbers in some periods would not affect the indoor temperature number SNtot under the proposed strategy decreased dramatically by
control (i.e. sufficient cooling can be provided). As listed in Table 5, 82% (194 vs. 35). With the conventional strategy, total energy con­
compared with the conventional strategy, the total switch number SNtot sumption can be reduced by 1.1% by sacrificing indoor thermal comfort
adopting the proposed strategy was decreased by 54% (63 vs. 29). The (i.e. indoor temperature cannot be controlled at the desired setpoint).
chiller mean stages in the test period adopting the proposed strategy (i.e. ζ0.5 of the conventional chiller sequencing control was 5.6 h, which was
2.55) were also less than that of the conventional strategy (i.e. 2.84), reduced to 0.18 h by the proposed strategy, decreasing by 96.8%.
indicating the potential energy saving. Compared with the conventional Fig. 13 presents the quantified risks in the decision-making process
strategy, the total energy consumed by the proposed strategy was when adopting the proposed strategy. Compared with Case 1, even the
reduced by 2.8%, without sacrificing indoor thermal comfort (i.e. ζ0.5 settings of senor bias were different, risk-based decision making can
was the same in two strategies) (see Fig. 11). operate the chillers while maintaining both energy-efficiency and robust
Fig. 10 presents the quantified risks in the decision-making process controls, as the optimal switching points can always be identified under
when adopting the proposed strategy. When the quantified risk oper­ measurement uncertainties and different operating conditions.
ating under N (i.e. current operating number) chillers (Rf, N) is higher
than βon (0.3), an additional chiller would be switched on. When there is 7. Conclusions
no risk of switching off a chiller (i.e., Rf, N-1 = 0), one chiller would be
staged off. The decision making is conducted based on risk assessment, A risk-based chiller sequencing control strategy was proposed to
offering a promising means for engineers to exploit the potential bene­ improve the robustness and energy efficiency of chillers in operation. As
fits/flexibility of chiller sequencing control. the core of the control strategy, an online stochastic decision-making
scheme was developed to optimize chiller staging. The risks of switch­
ing on/ off a chiller were evaluated based on the probabilistic fused
6.2. Results of Case 2: measured cooling loads < actual cooling loads cooling load and the probabilistic chiller maximum cooling capacity.
The cooling load and cooling capacity estimation models were cali­
Fig. 12A presents the variation in the number of the operating brated using the Bayesian Markov Chain Monte Carlo method, consid­
chillers adopting the conventional and proposed strategies in Case 2, ering the measurement uncertainties. Based on the results of the tests
where the measured cooling loads were lower than actual cooling loads.

Table 5
Control performance of proposed and conventional strategies in Case 1.
Strategy SNtot Nmean ECchi (kWh) ECpum (kWh) ECct (kWh) ECtot (kWh) ζ0.5 (h)
5 5 4 5
Conventional 63 2.84 5.00 × 10 1.4 × 10 7.88 × 10 7.20 × 10 0.18
Proposed 29 2.55 4.99 × 105 1.26 × 105 7.54 × 104 7.00 × 105 0.18

10
C. Zhuang et al. Applied Energy 280 (2020) 115983

Fig. 11. Quantified risks in the decision-making process of Case 1.

Fig. 12. Control performance of the proposed and conventional strategies in Case 2: (A) chiller operating number (B) indoor temperature of the thermal building.

Table 6
Control performance of proposed and conventional strategies in Case 2.
Strategy SNtot Nmean ECchi (kWh) ECpum (kWh) ECct (kWh) ECtot (kWh) ζ0.5 (h)

Conventional 194 2.23 5.07 × 105 1.10 × 105 7.30 × 104 6.90 × 105 5.60
Proposed 35 2.53 4.98 × 105 1.25 × 105 7.53 × 104 6.98 × 105 0.18

Fig. 13. Quantified risks in the decision-making process in Case 2.

11
C. Zhuang et al. Applied Energy 280 (2020) 115983

and implementation, some detailed conclusions can be drawn: [7] Yik FWH, Chiu TW. Measuring instruments in chiller plants and uncertainties in
performance evaluation. HKIE Trans 1998;5:95–9.
[8] Sun YJ, Wang SW, Huang GS. Chiller sequencing control with enhanced robustness
• The proposed strategy can dramatically decrease unnecessary chiller for energy efficient operation. Energy Build 2009;41:1246–55.
switching, which commonly occurs in conventional chiller [9] Huang GS, Sun YJ, Li P. Fusion of redundant measurements for enhancing the
sequencing control due to measurement uncertainties and changes in reliability of total cooling load based chiller sequencing control. Autom Constr
2011;20:789–98.
operating conditions. Compared with the conventional strategy, the [10] Liao YD, Huang GS, Ding YF, Wu HJ, Feng ZB. Robustness enhancement for chiller
switching frequency was decreased by more than 54%. The life span sequencing control under uncertainty. Appl Therm Eng 2018;141:811–8.
of chillers can therefore be prolonged with reduced maintenance [11] Yu FW, Chan KT. Experimental determination of the energy efficiency of an air-
cooled chiller under part load conditions. Energy 2005;30:1747–58.
costs. [12] Liu Z, Tan H, Luo D, Yu G, Li J, Li Z. Optimal chiller sequencing control in an office
• The probabilistic simplified physical models calibrated using the building considering the variation of chiller maximum cooling capacity. Energy
Bayesian Markov Chain Monte Carlo method can aggregate mea­ Build 2017;140:430–42.
[13] Shan K, Wang SW, Gao DC, Xiao F. Development and validation of an effective and
surement uncertainties and determine the profiles/distributions of robust chiller sequence control strategy using data-driven models. Autom Constr
cooling load and cooling capacity, which can be used for online risk 2016;65:78–85.
quantification reliably. [14] Zhuang CQ, Wang SW, Shan K. Probabilistic optimal design of cleanroom air-
conditioning systems facilitating optimal ventilation control under uncertainties.
• The proposed online stochastic decision-making scheme can effec­ Appl Energy 2019;253:113576.
tively optimize chiller staging, to achieve energy-efficient operation [15] Goel V, Grossmann IE. A stochastic programming approach to planning of offshore
of chillers and reliable thermal comfort control by providing suffi­ gas field developments under uncertainty in reserves. Comput Chem Eng 2004;28:
1409–29.
cient cooling. Results showed that the total energy use of the central
[16] –>Hemmati R, Saboori H, Saboori S. Stochastic risk-averse coordinated scheduling
cooling system could be reduced by 2.8% without sacrificing thermal of grid integrated energy storage units in transmission constrained wind-thermal
comfort. systems within a conditional value-at-risk framework. Energy 2016;113:762–75.
[17] Mehdizadeh A, Taghizadegan N, Salehi J. Risk-based energy management of
renewable-based microgrid using information gap decision theory in the presence
It is worth noticing that, in this study, the performance of the pro­ of peak load management. Appl Energy 2018;211:617–30.
posed strategy was evaluated under predefined risk thresholds. The [18] Westermann P, Evins R. Surrogate modelling for sustainable building design–A
performance of the strategy under different risk thresholds should be review. Energy Build 2019;198:170–86.
[19] Zhuang CQ, Wang SW. Risk-based online robust optimal control of air-conditioning
further tested to assess their impacts on the energy efficiency and systems for buildings requiring strict humidity control considering measurement
thermal comfort. uncertainties. Appl Energy 2020.
[20] Cano EL, Moguerza JM, Alonso-Ayuso A. A multi-stage stochastic optimization
model for energy systems planning and risk management. Energy Build 2016;110:
CRediT authorship contribution statement 49–56.
[21] Chou J-S, Ongkowijoyo CS. Risk-based group decision making regarding renewable
Chaoqun Zhuang: Methodology, Data curation, Validation, Soft­ energy schemes using a stochastic graphical matrix model. Autom Constr 2014;37:
98–109.
ware, Writing - original draft. Shengwei Wang: Conceptualization, [22] Mavromatidis G, Orehounig K, Carmeliet J. Comparison of alternative decision-
Writing - review & editing, Supervision, Funding acquisition. Kui Shan: making criteria in a two-stage stochastic program for the design of distributed
Resources. energy systems under uncertainty. Energy 2018;156:709–24.
[23] Li ZW, Huang GS, Sun YJ. Stochastic chiller sequencing control. Energy Build 2014;
84:203–13.
Declaration of Competing Interest [24] Gamerman D, Lopes HF. Markov chain Monte Carlo: stochastic simulation for
Bayesian inference. London: Chapman & Hall/CRC; 2006.
None. [25] Billera LJ, Diaconis P. A geometric interpretation of the Metropolis-Hastings
algorithm. Stat Sci 2001:335–9.
[26] Huang P, Augenbroe G, Huang G, Sun Y. Investigation of maximum cooling loss in
Acknowledgement a piping network using Bayesian Markov Chain Monte Carlo method. J Build
Perform Simul 2019;12:117–32.
[27] Min YR, Chen Y, Yang HX. A statistical modeling approach on the performance
The research presented in this paper is financially supported by a prediction of indirect evaporative cooling energy recovery systems. Appl Energy
general research grant (152075/19E) from the Hong Kong Research 2019;255:113832.
Grant Council (RGC). [28] Klein SA, Beckman WA, Mitchell JW, Duffie JA, Duffie NA, TLF, Mitchell JC, et al.
TRNSYS 18, a transient simulation program. University of Wisconsin-Madison,
USA: Solar Energy Laboratory; 2017.
References [29] Wang SW. Dynamic simulation of a building central chilling system and evaluation
of EMCS on-line control strategies. Build Environ 1998;33:1–20.
[1] Yu FW, Chan KT. Improved energy management of chiller systems by multivariate [30] Ma ZJ, Wang SW. Supervisory and optimal control of central chiller plants using
and data envelopment analyses. Appl Energy 2012;92:168–74. simplified adaptive models and genetic algorithm. Appl Energy 2011;88:198–211.
[2] Thangavelu SR, Myat A, Khambadkone A. Energy optimization methodology of [31] HK-BEAM Society. Hong Kong Building Environmental Assessment Method: HK-
multi-chiller plant in commercial buildings. Energy 2017;123:64–76. BEAM Version 5/04 Existing Buildings. HK-BEAM Society; 2004. https://www.
[3] Liao YD. Uncertainty analysis for chiller sequencing control. Energy Build 2014;85: beamsociety.org.hk/files/5-04%20(full)%2021feb05.pdf (last viewed date: Sept
187–98. 15, 2020).
[4] Huang S, Zuo WD, Sohn MD. Amelioration of the cooling load based chiller [32] Marquardt DW. An algorithm for least-squares estimation of nonlinear parameters.
sequencing control. Appl Energy 2016;168:204–15. J Soc Ind Appl Math 1963;11:431–41.
[5] Sun YJ, Wang SW, Xiao F. In situ performance comparison and evaluation of three [33] Li Q, Augenbroe G, Brown J. Assessment of linear emulators in lightweight
chiller sequencing control strategies in a super high-rise building. Energy Build Bayesian calibration of dynamic building energy models for parameter estimation
2013;61:333–43. and performance prediction. Energy Build 2016;124:194–202.
[6] Wang SW, Cui JT. Sensor-fault detection, diagnosis and estimation for centrifugal [34] Heo Y, Choudhary R, Augenbroe GA. Calibration of building energy models for
chiller systems using principal-component analysis method. Appl Energy 2005;82: retrofit analysis under uncertainty. Energy Build 2012;47:550–60.
197–213.

12