Yang_68_2021

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Applied Energy 288 (2021) 116648

Contents lists available at ScienceDirect

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

Experiment study of machine-learning-based approximate model predictive


control for energy-efficient building control
Shiyu Yang a, b, Man Pun Wan a, *, Wanyu Chen a, Bing Feng Ng a, Swapnil Dubey b
a
School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore 639798, Singapore
b
Energy Research Institute at NTU (ERI@N), Nanyang Technological University, Singapore 637553, Singapore

H I G H L I G H T S

• Proposed an approximate model predictive control for building energy management.


• The proposed control is based on a dynamic feedback machine learning model.
• The proposed control is experimentally implemented in two test buildings.
• The proposed control achieves up to 52% reduction of cooling energy.
• The proposed control is more than 100% faster than common model predictive control.

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

Keywords: The adoption of model predictive control (MPC) for building automation and control applications is challenged
Model Predictive Control (MPC) by the high hardware and software requirements to solve its optimization problem. This study proposes an
Machine-learning (ML) approximate MPC that mimics the dynamic behaviours of MPC using the recurrent neural network with a
Recurrent Neural Network (RNN)
structure of nonlinear autoregressive network with exogenous inputs. The approximate MPC is developed by
Air Conditioning and Mechanical Ventilation
(ACMV)
learning from the measured operation data of buildings controlled by MPC, therefore it can produce MPC-like
control for buildings without needing to solve the optimization problem, significantly reducing the computa­
tion load as compared to MPC. The proposed approximate MPC is implemented in two testbeds, an office and a
lecture theatre, to control the air-conditioning systems. The control performance of the approximate MPC is
compared to MPC as well as the original reactive control of the two testbeds. The approximate MPC retained
most of the energy and thermal comfort performance of MPC in both testbeds. For the office, the MPC and
approximate MPC reduced 58.5% and 51.6% of cooling energy consumption, respectively, as compared to the
original control. For the lecture theatre, the MPC and approximate MPC reduced 36.7% and 36.2% of cooling
energy consumption, respectively, as compared to the original control. Meanwhile, both approximate MPC and
MPC significantly improved indoor thermal comfort in the two testbeds as compared to their original control.
Despite having minor degradation in control performance the approximate MPC was more than 100 times faster
than MPC in generating optimal control commands in each time step.

can significantly improve building energy efficiency as suggested by


many studies [45] and one of the most promising control techniques is
1. Introduction
model predictive control (MPC) [6]. MPC could make use of a model of
the controlled building and takes weather conditions (e.g., air temper­
Buildings are responsible for 30% of global final energy consumption
ature, humidity, and solar radiation), internal heat loads (e.g., heat loads
[1] and more than 50% of building energy is consumed by heating,
from occupancy, equipment and lighting) and constraints (e.g.,
ventilation, and air-conditioning (HVAC) systems [2]. In tropical
comfortable ranges of indoor conditions and cooling capacity of HVAC/
countries, air-conditioning and mechanical ventilation (ACMV) systems
ACMV systems) into consideration for forward prediction and optimi­
also account for most, up to 70%, of the energy consumption in buildings
zation of building operation [7]. MPC has been demonstrated capable of
[3]. Using advanced control strategies for HAVC/ACMV management

* Corresponding author.
E-mail address: mpwan@ntu.edu.sg (M.P. Wan).

https://doi.org/10.1016/j.apenergy.2021.116648
Received 3 November 2020; Received in revised form 14 January 2021; Accepted 28 January 2021
Available online 22 February 2021
0306-2619/© 2021 Elsevier Ltd. All rights reserved.
S. Yang et al. Applied Energy 288 (2021) 116648

Nomenclature Symbols
PMV
̂ Normalized PMV
Abbreviations
Q
̂ Normalized Cooling Power
ACMV Air-Conditioning and Mechanical Ventilation ṁ Mass flow rate (kg/s)
AHU Air-Handling Unit
C Specific heat capacity (J/(K kg))
ANN Artificial Neural Network f() Function
ARX AutoRegressive network with eXogenous inputs
hod Hour of day
BAC Building Automation and Control J Objective
BMS Building Management System
N Prediction Horizon
CART Classification and Regression Trees n Number
COP Coefficient Of Performance
os Operation status (1 – on, 0 - off)
DDC Direct Digital Control Q heat flow rate (W)
FCU Fan Coil Unit
RH RH (%)
GLM Generalized Linear Model T Temperature (K)
HVAC Heating, Ventilation, and Air-Conditioning
u Input
LED Light-Emitting Diode W Weighting factor
LM Levenberg-Marquardt
y Output
LT Lecture Theatre ε Slack Variable
ML Machine Learning
MPC Model Predictive Control Subscripts
NARX Nonlinear AutoRegressive Network with eXogenous inputs AHU AHU
NTU Nanyang Technological University c cooling
PID Proportional–Integral–Derivative cw Chilled water
PLC Programmable Logic Controller cwr Chilled water return
PMV Predicted Mean Vote cws Chilled water supply
RC Resistance-Capacitance g Globe
RH Relative Humidity k Index variable
RNN Recurrent Neural Network o Outdoor
RT Regression Trees occ Occupancy
RTP Research Techno Plaza PMV PMV
SVM Support Vector Machine sa Supply air
TNDD Time-Delay Neural Networks t Time
VAV Variable Air Volume VAV VAV
VSD Variable Speed Drive z Zone

achieving substantial energy savings and significantly improved human state-space building model for MPC, which was implemented in an office
comfort for buildings, as compared to conventional reactive control (e. [19] and a lecture theatre [20] for real-time building optimization and
g., proportional–integral–derivative (PID)), in numerous previous control. However, the simplified building models still need to retain a
studies [8]. certain complexity (e.g., one state to represent one room) to capture the
However, several technical challenges are still limiting MPC from dynamics of every room. Therefore, despite the adoption of the simpli­
wider adoption for building automation and control (BAC) [9]. One key fied building modelling approach, MPC is still computationally expen­
challenge is that the optimization problem of MPC is computationally sive for large buildings with many rooms.
expensive to solve [10], making it prohibitive to implement MPC Another major hurdle for MPC adoption is that MPC requires so­
through typical industrial controllers, such as programmable logic phisticated optimization tools/libraries to solve the optimization prob­
controller (PLC) or direct digital controller (DDC), for real-time building lem [11]. These tools/libraries are usually programmed in high-level
control and optimization applications [11]. Much effort has been made languages (e.g., Python or MATLAB) and are not supported by most
to reduce the computation load of MPC, especially on the development current BAC hardware and software [21]. One possible solution to
of simplified, usually linear, building models [1213] for computational overcome these two challenges is to approximate the control laws of
efficiency. Bianchini et al. [14] developed a linear building thermal MPC by learning from the operation data of a building when it is
model based on AutoRegressive network with eXogenous inputs (ARX) controlled by MPC using machine learning (ML) techniques such as
and trained the model using the simulation data of a detailed building artificial neural network (ANN). The approximate MPC control laws (or
model. The linear building model was used to develop a MPC for con­ ‘Approximate MPC’) is essentially a data-driven model approximating
trolling heating systems. Hazyuk et al. [15] developed a low-order the correlations between the input and output variables of MPC. Thus,
building thermal model by thermal resistance–capacitance (RC) approximate MPC using ML learning methods (e.g., ANN) could be
network with parameters identified using the simulation data of a explicitly written in mathematical expressions that could be easily
detailed building model. The low-order building thermal model was implemented in industrial low-power control hardware (e.g., PLC or
used to construct a MPC controller for a heating system to optimize DDC) and low-level programming languages (e.g., structured text).
indoor thermal comfort and heating energy consumption [16]. Picard Some previous studies applied ML or statistical models on the
et al. [17] converted a high-order building model into reduced-order operation data to identify decision rules from MPC controllers to
models using the balanced truncation method for constructing MPC develop approximate MPC controllers, which were subsequently used to
for a heating system. Their simulations show that MPC using reduced- replace the original MPC controllers for building control. May-
order models requires much less computational effort compared to Ostendorp et al. [2223] used various data mining techniques
using a high-order building model. Yang et al. [18] developed a linear including generalized linear model (GLM), classification and regression

2
S. Yang et al. Applied Energy 288 (2021) 116648

trees (CART) and adaptive boosting to identify decision rules from the MPC.
simulated operation data of a MPC-controlled window. They applied the
similar technique to identified decision rules of a MPC-controlled water 2. Methodology
pump in a cooling system [24]. Domahidi et al. [25] used the adaptive
boosting method to identify decision rules from the simulated operation 2.1. Overview
data of a MPC-controlled HVAC system. Le et al. [26] used the support
vector machine (SVM) method to identify decision rules from simulated Fig. 1 shows an overview of the methodology used in this study. It
operation data of a MPC-controlled window blind. The identified control starts with having a MPC controller (consisting of a building model and
rules were programmed into simplified controllers to replace the orig­ optimization algorithm) to control the testbed building for a period of
inal MPC controllers in these studies. Despite having demonstrated the time. Building operation data collected during this period are to be used
feasibility of approximate MPC, these previous attempts make use of as the input/output data for training the approximate MPC controller.
decision rules that only generate control commands in a discrete deci­ The input data are the input variables of the MPC controller, including
sion space (e.g., choosing between on/off, open/closed or from a the measured disturbances (e.g., weather and internal heat loads) and
discrete set of values), missing the essential continuous control behavior room conditions (e.g., PMV, air temperature, globe temperature and
(choosing from a continuous decision space) of MPC [27] for controlling humidity). The output data include the cooling power set point of the
modulating actuators (e.g., water valves, water pumps and fans) of ACMV system, which is the control command generated by the MPC
HVAC/ACMV systems. controller. The approximate MPC controller is based on the NARX RNN,
A few simulation studies investigated the direct approximation of the a dynamic feedback ML model that approximates the dynamic correla­
continuous control laws of MPC using ML or statistical modelling for tions between the input variables and the control command of the MPC
BAC applications. Klauco et al. [28] developed an approximate MPC controller. Thus, the approximate MPC controller can perform MPC-like
controller by employing regression trees (RT) for controlling an HVAC control of a building with reference to the measured disturbances, room
system. Žáčeková et al. [29] developed an approximate MPC using conditions and the control command at the last time step, without
nonlinear regression for controlling an HVAC system. Robillart et al. needing to solve the optimization problem. The approximate MPC
[30] developed an approximate MPC controller using beta regression for controllers are then implemented in the testbed building to take over the
controlling a heating system to shift the heating load. Drgoňa et al. [31] control from the original MPC controllers. Control performance of the
developed a few approximate MPC controllers by using the time-delay approximate MPC controllers in the testbed building is experimentally
neural networks (TDNN) and RT for controlling an HVAC system. evaluated and compared against the original MPC controllers.
These studies have demonstrated the feasibility of approximate MPC to
mimic the continuous control behavior of MPC. However, the main
2.2. Testbeds
research gap of these previous studies is that they only focused on using
feed-forward ML/statistical models to develop approximate MPC. Thus,
Experiments were conducted in two testbeds including (1) an office
approximate MPC employing feedback model, which is superior to feed-
at level 3 of the Research Techno Plaza building representing typical
forward ML/statistical models in capturing dynamic behaviors [32], is
commercial built environments with moderate occupancy load and (2) a
rarely studied for BAC applications. Experimental evidence to support
lecture theatre (LT) on the campus of Nanyang Technological University
the feasibility and performance of approximate MPC is also lacking in
(NTU) in Singapore representing high-heat-load built environments
the literature.
with heavy, high-dynamic occupancy.
To fill these research gaps, this study proposes an approximate MPC
system that approximates the continuous control laws of a MPC for
2.2.1. Office testbed
optimizing building energy efficiency and thermal comfort. The pro­
Fig. 2 shows the internal layout and setup of the office testbed (or
posed approximate MPC features a dynamic feedback ML model, a
‘the office’). The office is surrounded by adjacent air-conditioned spaces
recurrent neural network (RNN) with a Nonlinear AutoRegressive
including other offices, a corridor and an atrium, thus, the weather does
network with eXogenous inputs (NARX) structure, which forms the main
not affect the internal thermal conditions of the office through the fa­
novelty of this study. The control performance of the proposed
cades. The floor area of the office is 46 m2 and the floor-to-ceiling height
approximate MPC system was investigated experimentally. Two
is 3 m. The major internal heat gains of the office are from the occupants,
approximate MPC controllers were developed and experimentally
computers and light-emitting diode (LED) lighting fixtures. Conditioned
implemented for real-time control of the ACMV systems in an office
air from a fan coil unit (FCU), which also serves a few adjacent spaces, is
testbed and a lecture theatre (LT) testbed. The building operation data
supplied into the office through a variable air volume (VAV) box, which
used to train the approximate MPC controllers were obtained from
has a maximum airflow capacity of 1,320 m3/h. The VAV box supplies
measurements made in the two testbeds when they were controlled by
the conditioned air through three 4-way diffusers as shown in Fig. 2. The
the MPC controllers developed in a previous study [33]. The control
VAV box is originally controlled by a thermostat, which regulates the
performance of the approximate MPC, in terms of energy efficiency and
supply air flowrate through an air damper according to a room air
indoor thermal comfort, was evaluated and compared to the original
temperature set point of 23 ◦ C. The ACMV system of the office operates

Fig. 1. An overview of the methodology used to develop the approximate MPC and performance evaluation.

3
S. Yang et al. Applied Energy 288 (2021) 116648

Fig. 2. Internal layout and setup of the office [33].

during 8 am – 6 pm of workdays and is switched off for the rest of the 2.2.2. Lecture theatre testbed
time. The LT testbed (or ‘LT’) has a seating capacity of 240 and a floor area
For the implementation of MPC and approximate MPC, the office is of 255 m2. The major internal loads are the heat gains from the occu­
installed with additional sensors including a combined temperature-RH pants and internal lighting. The whole room space is considered as one
(relative humidity) sensor and a thermal comfort measurement device, thermal zone, which is air-conditioned by an air-handling unit (AHU)
which measures air velocity, temperature, RH and globe temperature for that consists of a cooling coil and a variable speed drive (VSD) fan as
evaluation of predicted mean vote (PMV) index [34]. Occupant’s shown in Fig. 3. The airflow volume capacity of the AHU is 14,882 m3/h.
metabolic rate and clothing level are assumed to be constant at 1.2 met Chilled water is supplied from a central chiller plant that serves the
and 0.57 clo in PMV calculation, according to the typical activity level cluster of buildings where LT is located. The original building man­
and clothing of the occupants in the office. Another combined agement system (BMS) of LT regulates the fan speed according to the
temperature-RH sensor and an airflow sensor are installed in the supply room air temperature set point of 24 ◦ C and the opening of the chilled
air duct leading to the VAV box to measure the conditions of the supply water valve (WV) according to a varying supply air temperature set
air. Sensible cooling power supplied to the office through the VAV box at point, as described in Yang et al. [20]. When the control is switched to
any given point of time is calculated based on the measured difference in MPC, the AHU fan speed is regulated according to a supply air tem­
supply air and room air temperatures as well as the measured supply air perature set point of 15 ◦ C and the chilled water valve is regulated to
flowrate. More details of the office, its ACMV system and instrumenta­ chase the cooling power set points generated by the MPC controller. The
tion are described by Yang et al. [33]. approximate MPC controller shares the similar control approach as the

Fig. 3. A schematic diagram of the ACMV system serving LT [33].

4
S. Yang et al. Applied Energy 288 (2021) 116648

MPC controller with the exception that the cooling power set point is will unnecessarily increase the computation load in solving the opti­
generated by the approximate MPC algorithm instead of the MPC mization problem of MPC [37]. The output variable (manipulated var­
algorithm. iable) from the MPC is the optimal cooling power set point for the ACMV
For the implementation of MPC and approximate MPC, additional systems, i.e., the VAV box in the office or the AHU in LT. The input
sensors are installed in LT, as shown in Fig. 3, for measuring indoor and variables of the MPC include measured disturbances, feedback of
outdoor conditions as well as ACMV variables. The indoor thermal measured states as well as feedback of the measured manipulated vari­
comfort is measured by the PMV index which is determined from the able. More details of the MPC could be found in the previous study [33].
measured globe temperature, air temperature, air velocity and RH by A MPC system was developed and implemented in both testbeds
sensors mounted at the seat near the middle of LT. Occupant’s metabolic previously [33]. To obtain training data (input/output) for later devel­
rate and clothing level are assumed to be constant at 1.2 met and 0.5 clo, opment of approximate MPC, the MPC system controlled the VAV box in
respectively, according to the activity levels and clothing of the occu­ the office for 20 workdays (February 2019 – March 2019) and controlled
pants in LT. Cooling power provided to LT at any point of time is the AHU for 10 days with classes (9th – 15th September 2019 and 23rd –
measured by a thermal energy meter installed across the AHU cooling 29th September 2019). Building operation data were recorded at 5-min­
coil, measuring the chilled water mass flowrate, chilled water supply ute intervals. Further details of the MPC system and the experimental
and return temperatures. The fan power is measured by an electric implementation in the two testbeds can be found in Yang et al. [33].
power meter. The occupancy level in LT is estimated based on the
steady-state CO2 balance [35] in LT using the measured CO2 concen­
2.4. Approximate MPC system
trations in supply air and return air. More details of LT, ACMV system
and instrumentation are described by Yang et al. [2033].
The proposed approximate MPC system consists of an approximate
MPC controller and a local controller as illustrated in Fig. 4. With the
same measured input variables of MPC as described in Section 2.3
2.3. MPC**a
including measured disturbances (e.g., weather conditions and internal
heat loads), feedbacks of measured states (e.g., PMV, air temperature,
MPC controllers developed in a previous study [33] are employed as
globe temperature, humidity), as well as the feedback of measured
the state-of-art ‘teacher’ for the approximate MPC controllers developed
manipulated variable (cooling power set point), the approximate MPC
in this study. Instead of chasing a set point air temperature as the
controller generates a near-optimal cooling power set point for the next
original thermostat control does in the office or the original BMS does in
time step and sends to the local controller. The local controller, which is
LT, the objectives of the MPC are, instead, to minimize cooling energy
the same as the one in the MPC system [33], chases the cooling power set
consumption and deviation of indoor PMV from thermal neutrality
point received by regulating the actuators of the ACMV systems (air
(PMV = 0), as shown by the cost function described by Eq. (1).
damper in VAV box or chilled water valve WV in AHU) based on the PID
( )
∑ N ∑
N ( ) scheme. By ‘learning’ from the MPC as described in Section 2.3, the
J = Minimize ̂2
Q + [WPMV
̂
PMV t+k|t ] 2
+ W ε
ε t+k|t (1) approximate MPC follows the objectives of the MPC implicitly, which
c,t+k|t
are to minimize cooling energy consumption and minimize the deviation
k=1 k=1

of indoor PMV from thermal neutrality.


where symbols J, Q, ̂ PMV,
̂ W, N, and ε refer to objective, normalized
The approximate MPC is essentially a ML model that can predict
cooling power, normalized PMV, weighting factor, prediction horizon,
MPC-like control commands (cooling power set point for ACMV) based
and slack variable. Subscript k, t, c and PMV refer to index variable, time,
on the same input variables of MPC. In this study, the RNN with NARX
cooling and PMV. Normalized cooling power is the ratio of instanta­
structure, as expressed by Eq. (2), is employed to approximate the dy­
neous cooling power to the cooling capacity of the ACMV system.
namic behaviors of MPC for developing a ML model to work as the
Normalized PMV is the ratio of the instantaneous PMV to the magnitude
approximate MPC. The NARX RNN is a dynamic network that is suitable
of the comfortable PMV range (-0.5 ≤ PMV ≤ 0.5) [36]. Particularly, the
for time-series predictions [3839] and, thus, suitable for approximating
three terms on the right-hand side of Eq. (1) refer to costs of cooling
dynamic behaviors. The sigmoidal activation function, commonly used
energy consumption, thermal discomfort and violation of constraints
for ANN [40], was employed in this study.
within the prediction horizon, respectively. WPMV is the weighting factor ( )
for indoor PMV to allow for tuning the relative weighting between yt = f yt− 1 , yt− 2 , ⋯yt− ny , ut− 1 , ut− 2 , ⋯ut− nu (2)
thermal comfort (second term) and cooling energy consumption (first
term) in the cost function. In this study, WPMV is set at 4, whereas the where symbols f(), y, u, t and n refer to function, output, input, time
weighting factor for the cooling energy consumption is 1. This setting and number.
prioritizes indoor PMV close to thermal neutrality (PMV = 0) over The NARX RNN was trained using the Levenberg-Marquardt (LM)
minimizing the cooling energy. Wε is the weighting factor for the cost of algorithm in MATLAB deep learning toolbox [41] and the building
constraint violations (third term) and its value is set at 10000. With this operation data measured during MPC control, as described in Section
setting, any violation of constraints would make this term much larger 2.3. One hidden layer is selected for the NARX RNN, which can provide
than the first two terms, i.e., very small tolerance to constraint violation sufficient performance for most modelling problems [42] and avoid the
[33]. possibility of overfitting due to having too many hidden layers [43]. The
The optimization problem described by Eq. (1) is constrained by the number of neurons in the hidden layer was determined by a ‘trial-and-
comfortable range of indoor PMV (− 0.5 ≤ PMV ≤ 0.5) [36] and the error’ method [44]. Multiple trials of RNN training are performed
cooling capacity of the ACMV systems. The control interval of the MPC is starting from a small hidden layer size (number of neurons) to a point
5 min and the prediction horizon is 12-control-interval long. The pre­ that training errors cannot be reduced further by enlarging the hidden
diction horizon is selected based on the response time of indoor thermal layer size. This final hidden layer size is used for the NARX RNN, which
conditions in the two testbeds. During the morning start-up of ACMV in becomes the approximate MPC controller. The same configuration of
both testbeds, it takes maximum 40 min for the indoor PMV to reach and NARX RNN (except for the hidden layer sizes and the input variables) is
stablise at the best thermal comfort [33], suggesting that 40 min would used for both the office and LT testbeds to approximate the control laws
be enough to cover the response time of the indoor thermal conditions of their respective MPC controllers. The hidden layer size of NARX RNN
for the two testbeds. The selected prediction horizon of 12 control in­ for each testbed is determined based on the above-described ‘trial-and-
tervals (60 min) is long enough to cover the response time for optimizing error’ method. The input variables of the MPC controllers of the two
building future response. Setting a too much longer prediction horizon testbeds are different due to the difference in the internal setup (e.g., the

5
S. Yang et al. Applied Energy 288 (2021) 116648

Fig. 4. A schematic diagram showing the architecture of the proposed approximate MPC system.

number of occupants in LT is much more dynamic than the office) and


Table 1
structure (e.g., LT is exposed to the outdoor environment while the office
Listed of input/output data measured from the office for training the approxi­
is not) of the two testbeds. Details of the input/output variables of the
mate MPC controller.
NARX RNNs for the two testbeds are given below in Section 2.4.1 and
2.4.2. Variables Description Variable type in MPC

Qc,VAV Cooling power set point for the VAV box Manipulated variable
2.4.1. Approximate MPC controller for the office hod Hour of day Measured disturbance
os ACMV operation status
The approximate MPC controller (NARX RNN model) for the office,
RHz Room air RH
trained with the measured operation data obtained when the office was Tz Room air temperature Measured state variable
controlled by MPC, is illustrated in Fig. 5. The input data include hour of
day (hod) obtained from the CPU timer of the control server, operation
status (os) of the ACMV system obtained from the ACMV schedule, temperature (representing return air temperature) and supply air tem­
measured room air RH (RHz) and measured room air temperature (Tz), perature were measured using the instrumentation described in Section
as listed in Table 1. Some disturbances are unmeasured as they could be 2.2.1.
represented by other measured disturbances or have relatively small The input/output data set collected, as described in Section 2.3, was
effects on room conditions as compared to other measured disturbances, divided into three subsets randomly picked across the data collection
reducing instrumentation requirement and computational complexity of period, 70% for model training, 15% for model validation and the
MPC. The disturbances of internal loads are unmeasured and are rep­ remaining 15% for testing. 20 neurons of the hidden layer are selected
resented by the measured disturbance os (ACMV is on when os is 1 and for the approximate MPC controller based on the above-mentioned
ACMV is off when os is 0) by assuming that the temporal variation of ‘trial-and-error’ method. The R-values of training, validation and
internal loads during a day follows the ACMV operation schedule (in­ testing phases are 0.993, 0.991 and 0.992, respectively, which indicate
ternal loads are at 0% when ACMV is off, internal loads are at 100% that the trained approximated MPC controller has high prediction ac­
when ACMV is on). The thermal conditions (room air temperature and curacy in mimicking the behaviors of the MPC controller for predicting
humidity) in the adjacent spaces are treated as unmeasured disturbances cooling power set point. The R-values of the validation and testing
as the adjacent spaces are also air-conditioned. Hence, the difference of phases are very similar to that of the training phase, suggesting no
the thermal conditions between the adjacent spaces and the office is obvious overfitting to the training data.
expected to be small. Heat transfer between the office and the outdoors
through the building envelope was not considered as the office is not 2.4.2. Approximate MPC controller for LT
directly exposed to the outdoor environment. The output data are the Fig. 6 shows the schematic diagram of the approximate MPC
cooling power set point (Qc) generated by the MPC for the VAV box. The controller for LT. The input data include ACMV operation status (os),
sensible cooling power of the VAV box in the office is defined as, measured outdoor air temperature (To), measured outdoor air RH (RHo),
estimated number of occupants (nocc), measured room air temperature
Qc,VAV = ṁsa Csa (Tz − Tsa ), (3) (Tz), measured room air RH (RHz) and measured room globe tempera­
ture (Tg) as listed in Table 2. Besides the number of occupants, which is
where ṁ, C, Q and T refer to mass flow rate (kg/s), specific heat capacity
highly dynamic in LT and is estimated based on measured CO2 con­
(J/(K kg)), heat flow rate (W) and temperature (K), respectively. Sub­
centrations in supply air and return air as described in Section 2.2.2, the
scripts c, sa, VAV, and z refer to cooling, supply air, VAV box and thermal
other internal heat load due to lighting power is treated as unmeasured
zone, respectively. The supply air mass flow rate, thermal zone air
disturbance and represented by os with the same consideration for that

Fig. 5. Schematic diagram of the approximate MPC controller for the office. Fig. 6. Schematic diagram of the approximate MPC controller for LT.

6
S. Yang et al. Applied Energy 288 (2021) 116648

Table 2 Thermostat/BMS cases using the measured data from the two testbeds.
List of the input/output data measured from LT for training the approximate The test results were measured at 5-minute intervals from the two
MPC controller. testbeds. The statistical distributions of indoor PMV are plotted and
Variables Description Variable type in MPC analyzed to evaluate the thermal comfort performance of the three cases
Qc,AHU Cooling power set point for AHU Manipulated variable
for each testbed. The energy performance of the three cases in the office
os ACMV operation status Measured disturbance testbed is compared using the measured cooling power of the VAV box.
To Outdoor air temperature The energy performance of the three cases in LT is compared using the
RHo Outdoor air RH measured energy consumption (cooling energy + supply fan electricity)
nocc Number of occupants
of the AHU.
Tz Room air temperature Measured state variable
RHz Room air RH
Tg Room globe temperature 3.1. Test results in the office

Fig. 7 shows the time series of measured indoor PMV and room air
of the office as described in Section 2.2.1. The thermal conditions in temperature of the three cases during office hours of the test period (20
adjacent spaces are also considered as unmeasured disturbances with days per case). The results include the median value and interquartile
the same consideration for that of the office. The heat load from outdoor range of these time-series to show control characteristics and the fluc­
solar radiation is not considered as LT is located within shading [20]. tuation of indoor conditions of different controllers. The room air tem­
The output data is the cooling power set point (Qc) generated by the MPC perature is measured by the combined temperature-RH sensor and the
controller for the AHU. The cooling power of the AHU in the office is indoor PMV is measured by the thermal comfort measurement device
defined as, with globe temperature, air velocity and combined temperature-RH
Qc,AHU = ṁcw Ccw (Tcwr − Tcws ) (4) sensors, as shown in Fig. 2. Fig. 7 (a) & (d) show that the original
thermostat overcooled the office during the morning start-up period (8
where subscripts AHU, cw, cws and cwr refer to AHU, chilled water, am – 9:30 am). This can be attributed to the reactive nature of the
chilled water supply and chilled water return. thermostat’s control logic control that overshot the room air tempera­
The input/output data set collected, as described in Section 2.3, in LT ture set point (23 ◦ C) as it was trying to cool down the room during
was also divided into three subsets in the same way as that did for the morning start-up. The morning overshoot issue was eliminated with the
office. 20 neurons of the hidden layer were selected for the approximate MPC, with its prediction and optimization capabilities, which cooled the
MPC controller based on the above-mentioned ‘trial-and-error’ method. room air temperature to near thermal neutrality and stayed around there
The R-values of training, validation and testing phases are 0.998, 0.997 for the rest of the day, as shown in Fig. 7 (b) & (e). Instead of chasing a
and 0.998, respectively. The consistency of R-values of the three phases set point reactively as the thermostat did, MPC proactively predicted the
is, again, suggesting no obvious overfitting to the training data. future indoor PMV using the built-in building model and sequentially
optimized indoor thermal comfort (minimizing the deviation of indoor
2.5. Experimental arrangements PMV from thermal neutrality) and ACMV energy efficiency (minimizing
cooling energy consumption) based on its objective function (Eq. (1)).
In each testbed, the approximate MPC controller was implemented to Hence, the MPC maintained indoor PMV on the positive side but close to
perform real-time control of the ACMV system (the VAV box in the of­ the thermal neutrality at most office hours as shown in Fig. 7 (b), to
fice, the AHU in LT). The approximate MPC controller was programmed avoid overcooling and maintain best thermal comfort. By mimicking the
and executed in the same control server that the MPC controller was. control laws of the MPC, the approximate MPC also prevented the
The control performance of the approximate MPC controller is morning overshoot as shown in Fig. 7 (c) & (f). Both the MPC and the
compared to the MPC controller and the original reactive base control approximate MPC kept indoor PMV and room air temperature within
(Thermostat in the office, BMS in LT). The results of the MPC controller much narrower ranges from thermal neutrality as compared to that of
and the original control are obtained from the previous study [33] for the thermostat did. It suggests that the MPC and the approximate MPC
this comparison. Time periods of the test cases are summarised in maintained much more stable indoor thermal conditions than the ther­
Table 3. In the test periods, the office was operating during 8:00 am – mostat did.
6:00 pm of workdays and LT was operating according to the class Fig. 8 summarises the measured indoor PMV and room air temper­
schedule listed in Table 4. ature during office hours using box-and-whisker plots. These plots show
the statistical distributions of the measured data (maximum, 75th
3. Results percentile, median, mean, 25th percentile, and minimum) obtained
from 20 days of measurements per case in the office. The same config­
To evaluate the control performance of the approximate MPC, the uration of the box and whisker plot is applied to Figs. 9–11. Fig. 8 (a)
test results of the Approximate MPC case is compared to the MPC and shows that all the three controllers were able to maintain indoor PMV
within the comfortable range (-0.5 ≤ PMV ≤ 0.5) for all the office hours.
This could be due to the fact that the office was surrounded by adjacent
Table 3
air conditioning spaces and, thus, away from direct impact of weather
Time periods of the test cases.
disturbances. Despite so the MPC and the approximate MPC still provide
Office LT
thermal comfort advantage over the reactive thermostat by maintaining
Case Test Period (20 days/ Case Test Period (10 the indoor PMV much closer to thermal neutrality. MPC maintained the
case) days/case) most stable indoor thermal comfort condition, as indicated by the nar­
Approximate April – May 2019 Approximate November 2019 rowest interquartile range of PMV, followed by approximate MPC and
MPC MPC then thermostat, echoing the observations from Fig. 7. The MPC and the
MPC [33] February – March MPC [33] August –
approximate MPC also maintained indoor PMV on the positive side for
2019 September 2019
Thermostat December 2018 – BMS [33] August – most time of the office hours, suggesting that the optimization for
[33] January 2019 September 2019 cooling power saving caused the MPC to incline towards positive PMV.
Control server: Intel Core i5-3317U CPU Control server: Intel Core i5-6200U This control characteristic was mimicked by the approximate MPC. In
and 4G RAM with MATLAB/Simulink CPU and 8 GB RAM with MATLAB/ contrast, without any consciousness to save cooling power, the ther­
platform Simulink platform
mostat maintained indoor PMV on the negative side for most time of the

7
S. Yang et al. Applied Energy 288 (2021) 116648

Table 4
Weekly class schedule of LT during the test period [33].

9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00 21:00
Monday
Tuesday
Wednesday
Thursday
Friday
Key: - class

Fig. 7. Time series of indoor conditions measured in the office during office hours of the whole test period (20 days per case). (a), (b) & (c) Room air temperature.
(d), (e) & (f) Indoor PMV.

Fig. 8. Statistical distributions of (a) indoor PMV and (b) room air temperature measured in the office during office hours of the whole test period.

office hours. The set-point-chasing nature of the thermostat can be seen temperature shown in Fig. 7 and Fig. 8 suggests that the approximate
more clearly in Fig. 8 (b). It shows that both the mean and median room MPC reproduced the control characteristics of MPC, in terms of con­
air temperature in the Thermostat case is 23.4 ◦ C, very close to the room trolling room thermal conditions, fairly closely.
air temperature set point of 23 ◦ C in the thermostat. Mean and median Fig. 9 summarises the statistical distributions of measured cooling
room air temperature in the MPC case is a higher 24.1 ◦ C, which is the power supplied by the VAV box to the office during office hours using
outcome of chasing the objectives (minimize cooling energy consump­ box-and-whisker plots. The cooling power is evaluated with the
tion and minimize the deviation between indoor PMV and thermal measured temperature and flowrate of supply air as well as room air
neutrality) defined in Eq. (1). Mean and median room air temperature in temperature, as described in Section 2.2.1. The mean cooling power of
the Approximate MPC case are 24.2 ◦ C and 24.1 ◦ C, respectively, very the Thermostat case was the highest among the three cases at 1.88 kWh.
close to that of the MPC. The measured indoor PMV and room air The MPC case shows a substantial reduction in mean cooling power at

8
S. Yang et al. Applied Energy 288 (2021) 116648

control. The approximate MPC was able to capture most of the control
characteristics of the MPC and retained most of the benefits.
While retaining most benefits of the MPC, the approximate MPC
greatly reduced the computation load as compared to the MPC. The
average CPU processing time needed to generate the cooling power set
point for the VAV box in the office in a control interval was 2.5 s in the
MPC controller but largely reduced to 0.016 s in the approximate MPC
controller. The much higher computation efficiency of the approximate
MPC makes it much more adaptable to low-power hardware (e.g., PLC
and DDC used in buildings) as compared to the MPC.

3.2. Test results of LT

Fig. 10 shows the statistical distributions of room air temperature


Fig. 9. Statistical distributions of measured cooling power supplied by the VAV and indoor PMV measured during class time using the box-and-whisker
box during office hours of the whole test period (10 days per case). plots. The room air temperature is measured by the air temperature
sensor and the indoor PMV is measured by the globe temperature, air
0.78 kWh, a 58.9% reduction compared to the Thermostat case. By temperature, air velocity and RH sensors as described in Section 2.2.2.
mimicking the control laws of MPC, the approximate MPC retained most The tendency to overcool LT by the reactive BMS can be obviously seen
of the energy-saving potential of MPC by achieving 51.6% reduction in in Fig. 10 (a), as indicated by a mean PMV of − 0.35. Indoor PMV in the
mean cooling power reduction compared to the Thermostat case. Since BMS case went lower than the lower bound of the comfortable range
cooling power in all three cases was measured at the same intervals for (PMV = − 0.5), causing occupants to suffer from thermal discomfort, for
the same amount of time, the mean cooling power directly reflects the about 12% of the class time. In contrast, both the MPC and the
cooling energy consumption over the test period. The results suggest approximate MPC kept indoor PMV within the comfortable range for all
that the MPC achieved significant cooling energy reduction while the class time, with a mean indoor PMV of 0.07 and 0.01, respectively. It
improving indoor thermal comfort compared to the original thermostat indicates that significant improvements in thermal comfort were ach­
ieved by the MPC and the approximate MPC, as compared to the BMS.

Fig. 10. Statistical distributions of (a) indoor PMV and (b) room air temperature measured in LT during class time of the whole test period.

Fig. 11. Energy performance of the BMS, MPC and approximate MPC cases in LT. (a) Statistical distributions of measured cooling power supplied by the AHU, (b)
average weekly ACMV electricity consumption.

9
S. Yang et al. Applied Energy 288 (2021) 116648

Similar to that observed in the office, without the consciousness to save and advanced optimization software for the entire building operational
cooling energy, the reactive BMS control tends to overcool LT and kept life. Since approximate MPC can run on typical industrial controllers and
indoor PMV to be negative for all the class time, suggesting excessive its control algorithm is essentially running as a black-box, approximate
cooling energy used in the BMS case. MPC presents a much smoother learning curve for facility managers as
Fig. 10 (b) shows that both the mean and median room air temper­ compared to MPC, which has highly explicit building model and opti­
ature with the BMS control are 24.1 ◦ C, which is close to the room air mization algorithm.
temperature set point of 24 ◦ C, indicating the set-point-chasing control
characteristic of the BMS. In contrast, both the MPC and the approxi­ 4.2. Applicability and limitations of approximate MPC
mate MPC kept indoor PMV close to the thermal neutrality and on the
positive side for the majority of time, preventing the overcooling issue. The proposed approximate MPC is developed based on a fully data-
This indicates the control characteristics of MPC in chasing the multiple driven approach, i.e., when the proposed approximate MPC is imple­
targets of thermal comfort and minimizing cooling power. The approx­ mented in another building, operation data of that particular building
imate MPC mimicked these control characteristics fairly closely. As an will be needed to retrain and update the approximate MPC (in terms of
outcome of the control characteristics of the MPC and the approximate input variables, hidden layer size and model parameters). However, the
MPC, the mean and median room air temperature was 25.2 ◦ C in the framework for model and control algorithm configuration (NARX RNN)
MPC case and 25.1 ◦ C in the approximate MPC case. The only attribute and the methodology for controller development should be generally
that the approximate MPC fell short compared to the MPC was that the applicable. This is demonstrated in the current paper that the approxi­
approximate MPC could not keep the room conditions as stable as the mate MPC controllers in the two different testbeds share an identical
MPC did. This could be due to the fact that LT was directly subjected to model configuration framework and development process, as described
the impacts of weather (walls are directly exposed to outdoor) and in Section 2.1. The only differences are the hidden layer sizes and the
highly dynamic occupancy load. The dynamics of heat loads through input variables, as described in Section 2.4. The two chosen testbeds
walls and from occupancy might have added difficulties for the represent some common built environments. The office has moderate,
approximate MPC control algorithm. smooth occupancy load representing typical commercial built environ­
Fig. 11 summarizes the cooling energy consumption of the three ments. LT has high-dynamic and heavy occupancy load, representing
cases in LT, which is measured by the energy meter described in Section high-heat-density built environment.
2.2.2. Fig. 11 (a) shows the statistical distributions of measured cooling The experiments conducted in this study, to some extent, verified the
power supplied by the AHU to LT using box-and-whisker plots. Among resilience of the approximate MPC to stochastic disturbances, including
the three cases, the mean cooling power is the highest in the BMS case internal heat loads and weather conditions. The approximate MPC ran
with 32.1 kWh. The MPC and the Approximate MPC cases are signifi­ for 10 weekdays in November 2019 in LT, when LT was under stochastic
cantly lower in mean cooling power at 22.3 kWh and 22.1 kWh, occupancy and weather conditions. The estimated number of occupancy
respectively. Fig. 11 (b) shows the average weekly ACMV electricity varied between 0 and 180 with a median value of 66, depending on the
consumption of the three cases. ACMV electricity consumption consists actual class schedule running in LT. The sky clearness index [45] that
of electricity consumption by the AHU ‘Fan’ and ‘Cooling’, which is the represents weather conditions varied between 0.23 and 0.50 with a
thermal cooling power measured across the cooling coil divided by the median value of 0.36 and its distribution is shown in Table 5. Under the
estimated mean coefficient of performance (COP = 5.39) of the chiller stochastic disturbances, the approximate MPC still achieved most of the
serving LT. The MPC and approximate MPC achieved 36.7% and 36.2% benefits of the MPC, in terms of both energy saving and thermal comfort
reduction in ACMV electricity consumption as compared to the BMS. as discussed in Section 3.2. As the approximate MPC ‘learns’ from the
The approximate MPC has, again, demonstrated its ability to capture MPC, the resilience of the approximate MPC would continue to improve
most of the benefits of the MPC, in terms of energy saving and thermal if a stochastic or robust MPC is employed to generate the building
comfort. operation data to train the approximate MPC.
Similar to the CPU time results observed in the office, the approxi­ However, a limitation of the current study is that we could not verify
mate MPC also largely reduced the average CPU processing time needed the resilience of the approximate MPC for disturbances that go outside
to generate a cooling power set point for the AHU of LT in a control the range covered by the current experiments (internal load dynamics,
interval to 0.017 s, from 6.3 s in the MPC. tropical climate). Table 5 shows that the weather conditions of the
approximate MPC test period are representative of the intermediate sky
4. Discussion condition, which happens on 95% of days in a year in Singapore ac­
cording to the Singapore typical meteorological year (TMY) data [46].
4.1. Advantages of approximate MPC in practical implementation The resilience of the approximate MPC for the remaining 5% days with
overcast and clear sky conditions could be verified in future studies.
In the experiments, the proposed approximate MPC showed a few Another limitation of this study is that building heating was not
technical advantages over the MPC in practical implementation. The considered in the experiments as both testbeds are in tropics, where
approximate MPC significantly reduced the computational load, as heating is generally not needed in buildings. Adopting the proposed
compared to the MPC with only slightly compromised control perfor­ approximate MPC for regions where heating is needed would require
mance in terms of thermal comfort and energy saving compared to the measuring heating power data as one of the inputs for the development
MPC but still significantly better than the original reactive-based con­ and operation of the approximate MPC.
trols. The approximate MPC is also much easier to be implemented in
low-power control hardware (e.g., PLC and DDC) and low-level pro­ Table 5
gramming languages, as it does not need to solve optimization problems. Comparison of the statistical distributions of sky conditions between the
Thus, the approximate MPC is much more compatible with existing BMS Singapore TMY data and approximate MPC test period.
hardware and software than MPC, which requires high computing Ranges of sky Sky clearness Days of Singapore Days of approximate
power and advanced optimization software with dedicated libraries/ clearness index classification TMY data [46] MPC test period
tools. The development of approximate MPC adds to the initial devel­ 0.05–0.20 Overcast 14 0
opment of MPC since approximate MPC would have to ‘learn’ from MPC. 0.20–0.35 Intermediate 30 5
Despite needing additional effort at the initial development stage, 0.35–0.50 Intermediate 228 4
approximate MPC can run on low-power hardware for the rest of the 0.50–0.65 Intermediate 88 1
0.65–0.80 Clear 5 0
building’s operational life whereas MPC requires high-power hardware

10
S. Yang et al. Applied Energy 288 (2021) 116648

5. Conclusion [8] Serale G, Fiorentini M, Capozzoli A, Bernardini D, Bemporad A. Model predictive


control (MPC) for enhancing building and HVAC system energy efficiency: Problem
formulation, applications and opportunities. Energies 2018;11(3):631.
This paper proposed and developed an approximate MPC using the [9] Rockett P, Hathway EA. Model-predictive control for non-domestic buildings: a
recurrent neural network with a structure of nonlinear autoregressive critical review and prospects. Building Research & Information 2017;45(5):
network with exogenous inputs to approximate the control laws of MPC 556–71.
[10] Killian M, Kozek M. Ten questions concerning model predictive control for energy
for building automation and control applications. The approximate MPC efficient buildings. Build Environ 2016;105:403–12.
can mimic the behaviors of MPC in controlling buildings without [11] Cígler J, Gyalistras D, Široky J, Tiet V, Ferkl L. Beyond theory: the challenge of
needing to solve the optimization problem. The control performance of implementing model predictive control in buildings. Proceedings of 11th Rehva
world congress, Clima. Prague, Czech Republic. 2013.
the approximate MPC was evaluated through experiments in two test­ [12] Fontenot H, Dong B. Modeling and control of building-integrated microgrids for
beds, an office and a lecture theatre, to control their air conditioning optimal energy management–A review. Appl Energy 2019;254:113689.
systems. The control performance of the approximate MPC was [13] Privara S, Cigler J, Váňa Z, Oldewurtel F, Sagerschnig C, Žáčeková E. Building
modeling as a crucial part for building predictive control. Energy Build 2013;56:
compared to the original reactive controls of the two testbeds as well as 8–22.
MPC. In the office, the MPC and approximate MPC reduced 58.5% and [14] Bianchini G, Casini M, Vicino A, Zarrilli D. Demand-response in building heating
51.6% of cooling energy consumption, respectively, as compared to the systems: A Model Predictive Control approach. Appl Energy 2016;168:159–70.
[15] Hazyuk I, Ghiaus C, Penhouet D. Optimal temperature control of intermittently
original reactive control. In the lecture theatre, the MPC and approxi­ heated buildings using Model Predictive Control: Part I-Building modeling. Build
mate MPC reduced 36.7% and 36.2% of cooling energy consumption, Environ 2012;51:379–87.
respectively, as compared to the original reactive control. The energy [16] Hazyuk I, Ghiaus C, Penhouet D. Optimal temperature control of intermittently
heated buildings using Model Predictive Control: Part II–Control algorithm. Build
performance results suggest that the approximate MPC could retain most
Environ 2012;51:388–94.
of the energy-saving potential of MPC. The approximate MPC also [17] Picard D, Drgoňa J, Kvasnica M, Helsen L. Impact of the controller model
captured most of the thermal comfort performance of the MPC. The complexity on model predictive control performance for buildings. Energy Build
approximate MPC controllers were more than 100 times faster than the 2017;152:739–51.
[18] Yang S, Wan MP, Ng BF, Zhang T, Babu S, Zhang Z, et al. A state-space thermal
MPC controllers in both testbeds in determining the control commands model incorporating humidity and thermal comfort for model predictive control in
for each time step, suggesting that approximate MPC has much lower buildings. Energy Build 2018;170:25–39.
computational power requirements than MPC. This study demonstrated [19] Yang S, Wan MP, Ng BF, Dubey S, Henze GP, Rai SK, et al. Experimental study of a
model predictive control system for active chilled beam (ACB) air-conditioning
the feasibility of employing approximate MPC to retain much of the system. Energy Build 2019;203:109451.
benefits of MPC while largely lowering the control hardware computa­ [20] Yang S, Wan MP, Ng BF, Dubey S, Henze GP, Chen W, et al. Experimental study of
tion power demand for implementation. model predictive control for an air-conditioning system with dedicated outdoor air
system. Appl Energy 2020;257:113920.
[21] Huyck B, Ferreau HJ, Diehl M, De Brabanter J, Van Impe JF, De Moor, B, Logist F.
CRediT authorship contribution statement Towards online model predictive control on a programmable logic controller:
Practical considerations. Mathem Probl Eng. 2012.
[22] May-Ostendorp P, Henze GP, Corbin CD, Rajagopalan B, Felsmann C. Model-
Shiyu Yang: Conceptualization, Methodology, Software, Validation, predictive control of mixed-mode buildings with rule extraction. Build Environ
Investigation, Writing - original draft, Visualization, Project adminis­ 2011;46(2):428–37.
tration. Man Pun Wan: Conceptualization, Supervision, Writing - re­ [23] May-Ostendorp PT, Henze GP, Rajagopalan B, Corbin CD. Extraction of supervisory
building control rules from model predictive control of windows in a mixed mode
view & editing, Resources. Wanyu Chen: Formal analysis, building. J Build Perform Simul 2013;6(3):199–219.
Visualization. Bing Feng Ng: Supervision, Resources. Swapnil Dubey: [24] May-Ostendorp PT, Henze GP, Rajagopalan B, Kalz D. Experimental investigation
Resources. of model predictive control-based rules for a radiantly cooled office. HVAC&R
Research 2013;19(5):602–15.
[25] Domahidi A, Ullmann F, Morari M, Jones CN. Learning near-optimal decision rules
Declaration of Competing Interest for energy efficient building control. In: 51st IEEE Conference on Decision and
Control (CDC). USA: Hawaii; 2012. p. 7571–6.
[26] Le K, Bourdais R, Guéguen H. From hybrid model predictive control to logical
The authors declare that they have no known competing financial control for shading system: A support vector machine approach. Energy Build
interests or personal relationships that could have appeared to influence 2014;84:352–9.
the work reported in this paper. [27] Drgoňa J, Arroyo J, Figueroa IC, Blum D, Arendt K, Kim D, et al. All you need to
know about model predictive control for buildings. Annual Reviews in Control
2020.
Acknowledgement [28] Klaučo M, Drgoňa J, Kvasnica M, Di Cairano S. Building temperature control by
simple MPC-like feedback laws learned from closed-loop data. IFAC Proceedings
Volumes 2014;47(3):581–6.
This research is financially supported by JTC Corporation (contract [29] Žáčeková E, Pčolka M, Tabačk J, Těžký J, Robinett R, Čelikovský S, et al. In: July).
nos. N190107T00 and 2019-0607) and Smart Nation & Digital Gov­ Identification and energy efficient control for a building: Getting inspired by MPC.
ernment Office (SNDGO) of Singapore (Grant no. NRF2016IDM- IEEE; 2015. p. 1671–6.
[30] Robillart M, Schalbart P, Peuportier B. Derivation of simplified control rules from
TRANS001-031). an optimal strategy for electric heating in a residential building. J Build Perform
Simul 2018;11(3):294–308.
References [31] Drgoňa J, Picard D, Kvasnica M, Helsen L. Approximate model predictive building
control via machine learning. Appl Energy 2018;218:199–216.
[32] Zhang XS. Feedback Neural Networks. Neural Networks in Optimization.
[1] Abergel T, Dean B, Dulac J. Towards a zero-emission, efficient, and resilient
Nonconvex Optimization and Its Applications 2000;vol 46.
buildings and construction sector: Global Status Report 2017. Paris, France: UN
[33] Yang S, Wan MP, Chen W, Ng BF, Dubey S. Model predictive control with adaptive
Environment and International Energy Agency; 2017.
machine-learning-based model for building energy efficiency and comfort
[2] Razmara M, Maasoumy M, Shahbakhti M, Robinett III RD. Optimal exergy control
optimization. Appl Energy 2020;271:115147.
of building HVAC system. Appl Energy 2015;156:555–65.
[34] ISO. ISO 7730:2005 - Ergonomics of the thermal environment - Analytical
[3] Chua KJ, Chou SK, Yang WM, Yan J. Achieving better energy-efficient air
determination and interpretation of thermal comfort using calculation of the PMV
conditioning–a review of technologies and strategies. Appl Energy 2013;104:
and PPD indices and local thermal comfort criteria. Switzerland: ISO; 2005.
87–104.
[35] Wang S, Burnett J, Chong H. Experimental validation of CO2-based occupancy
[4] Shaikh PH, Nor NBM, Nallagownden P, Elamvazuthi I, Ibrahim T. A review on
detection for demand-controlled ventilation. Indoor Built Environ 1999;8(6):
optimized control systems for building energy and comfort management of smart
377–91.
sustainable buildings. Renew Sustain Energy Rev 2014;34:409–29.
[36] ASHRAE. Standard 55–2013: Thermal Environmental Conditions for Human
[5] Treado S, Chen Y. Saving building energy through advanced control strategies.
Occupancy. Atlanta, USA: ASHRAE; 2013.
Energies 2013;6(9):4769–85.
[37] Zhao S, Cajo Diaz RA, De Keyser R, Liu S, Ionescu CM. Nonlinear predictive control
[6] Afram A, Janabi-Sharifi F. Theory and applications of HVAC control systems–A
applied to steam/water loop in large scale ships. In: In 12th IFAC Symposium on
review of model predictive control (MPC). Build Environ 2014;72:343–55.
Dynamics and Control of Process Systems, including Biosystems; 2019. p. 868–73.
[7] Yang S, Wan MP, Chen W, Ng BF, Zhai D. An adaptive robust model predictive
control for indoor climate optimization and uncertainties handling in buildings.
Build Environ 2019;163:106326.

11
S. Yang et al. Applied Energy 288 (2021) 116648

[38] Çoruh S, Geyikçi F, Kılıç E, Çoruh U. The use of NARX neural network for modeling [42] Macukow B. In: Neural networks–state of art, brief history, basic models and
of adsorption of zinc ions using activated almond shell as a potential biosorbent. architecture. Cham: Springer; 2016. p. 3–14.
Bioresour Technol 2014;151:406–10. [43] Goodfellow I, Bengio Y, Courville A, Bengio Y. Deep learning 2016;Vol. 1:p. 2.
[39] Katić K, Li R, Verhaart J, Zeiler W. Neural network based predictive control of [44] Asadi E, da Silva MG, Antunes CH, Dias L, Glicksman L. Multi-objective
personalized heating systems. Energy Build 2018;174:199–213. optimization for building retrofit: A model using genetic algorithm and artificial
[40] Witten IH, Frank E, Hall MA, Pal CJ. Data Mining: Practical Machine Learning neural network and an application. Energy Build 2014;81:444–56.
Tools and Techniques. 4th Ed. San Francisco, CA, USA: Morgan Kaufmann; 2017. [45] Duffie JA, Beckman WA. Solar engineering of thermal processes. John Wiley &
[41] Beale MH, Hagan MT, Demuth HB. Deep learning toolbox™ reference. Natick, MA, Sons; 2013.
USA: The MathWorks; 2020. [46] ASHRAE, International Weather for Energy Calculations (IWEC Weather Files)
User’s Manual and CD-ROM, ASHRAE, 2001.

12

You might also like