Page 47 - IJAMD-2-1

International Journal of AI for
Materials and Design
ML-based MPC for multizone BAC

predictive capabilities and overall performance of building • Section 3 (MPC controller for ACMV): Details the
automation systems. design and implementation of the MPC controller for
Despite the promise of MPC, challenges remain in the ACMV system
accurately modeling the complex thermal dynamics • Section 4 (ML model development for MPC):
of buildings, which are influenced by varying weather Outlines the development of the ML models used
conditions, occupancy patterns, and building envelope for forecasting and their integration within the MPC
characteristics. In addition, the computational burden framework
associated with optimization across multiple zones • Section 5 (Results and discussion): Presents the
presents further difficulties. To address these challenges, performance evaluation, key findings, and a comparative
23
model-free approaches based on RL (such as Q-learning analysis with conventional control strategies and
24
and deep reinforcement learning ) have been explored, discusses the challenges encountered
25
with several studies demonstrating notable energy savings. • Section 6 (Conclusion): Summarizes the study’s
For instance, Ding et al. reported approximately 14% contributions, implications, and suggestions for future
26
energy savings using deep RL with inputs from EnergyPlus research.
simulations. However, the present study focuses on MPC 2. Testbed setup
due to its structured approach and its ability to directly
integrate system constraints into the optimization process. For the testbed, a commercial building located in Jurong
East, Singapore, was selected. The building is served by a
In this paper, we present a multizone MPC framework central chiller plant consisting of two chillers, each with
for the coordinated control of ACMV systems, aiming to a capacity of 438 refrigeration tons. The ACMV system in
optimize both occupant comfort and energy efficiency. To the building operates from 7 AM to 7 PM on weekdays.
evaluate performance, we compare the energy consumption The experiments were conducted in an 850 m multiuse
2
and thermal comfort achieved using our coordinated test space, partitioned into 11 zones, including two office
MPC approach with those obtained from a conventional spaces (OS1 – 2), six learning zones (LZ1 – 6), and three
reactive control strategy employing proportional-integral- common areas (CA1 – 3), as depicted in Figure 1. MPC
derivative (PID) control. The key novelties of our MPC was implemented to coordinate and control the ACMV
system are as follows: systems, as illustrated in Figure 2. The test space includes
(i) Large-scale, multizone implementation: This study operable partition walls between LZ1 and LZ2, as well as
demonstrates a real-world deployment of MPC in an among LZ4, LZ5, and LZ6. These partitioning walls can be
11-zone building (850 m ) with actual occupancy. This adjusted so that two zones (LZ1 and LZ2) or up to three
2
contrasts with previous studies, which were limited zones (LZ4, LZ5, and LZ6) can be merged into a single
to smaller-scale implementations (e.g., two zones ), zone. The open/closed state of these walls is determined by
27
simulated environments, or unoccupied buildings 29 proximity sensors, which output “1” if the walls are closed
28
(ii) Integration of ML-based weather forecasting: By (i.e., zones remain separate, as in LZ1 and LZ2) or “0” if the
incorporating ML-based weather forecasting into walls are open (zones merge into one larger room).
the MPC framework, our system proactively adjusts
control actions in anticipation of weather changes. The schematic diagram of the ACMV system installed
This represents an advancement over previous studies, in the test space is shown in Figure 2. The ACMV system
which typically relied on existing or historical weather serving the test space consists of two primary air-handling
data. 30-32 units (PAUs) for pre-conditioned fresh air supply and
16 fan coil units (FCUs) for cooling and dehumidification.
This work provides valuable insights into both The cooling coils in the PAUs and FCUs are supplied with
the challenges and benefits of implementing MPC in chilled water from the building’s central chiller plant
complex, real-world settings, particularly within tropical (not shown in Figure 2). Since the MPC system of this
environments. Consequently, it contributes significantly to study controls only the airside equipment of the ACMV
the advancement of scalable solutions for building energy system within the test space – a fraction of the entire
management. building – the net effect of MPC on the overall chiller plant
This paper is organized as follows: is not considered. Instead, cooling energy consumption,
• Section 1 (Introduction): Provides background, measured by British thermal unit (BTU) meters (or energy
motivation, and objectives of the study meters) installed at the cooling coils of the PAUs and FCUs,
• Section 2 (Testbed setup): Describes the heating, is used to evaluate the cooling energy consumed by the
ventilation, and air conditioning (HVAC) system and ACMV system in the test space. Each PAU is equipped with
sensor configuration of the testbed an on/off damper, a cooling coil, and a supply fan. The PAUs

Volume 2 Issue 1 (2025) 41 doi: 10.36922/ijamd.8161