Page 46 - IJAMD-2-1

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

1. Introduction datasets from real buildings or simulations, necessitating
extensive data collection. Hybrid models combine elements
In Singapore, owing to its hot and humid climate of both approaches, estimating certain physical parameters
throughout the year, commercial buildings typically through data-driven methods.
spend more than 60% of their electricity consumption on
air conditioning, mechanical, and ventilation (ACMV) Data-driven MPC strategies have demonstrated
systems. On March 31, 2020, Singapore submitted its considerable promise in achieving both optimal thermal
1
Long-Term Low Emissions Development Strategy to comfort and high energy efficiency. Over the years,
11
10
the United Nations Framework Convention on Climate researchers such as Yang et al., Yang and Wan have
Change, pledging to halve its emissions from their peak to enhanced MPC by simplifying building models through
12
33 MTCO e and to achieve net-zero emissions by 2050. At linearization techniques. Široký et al. reported energy
2
2
present, the Building Construction Authority (BCA) has savings of 15 – 30% in a university building by integrating
reported that high-performance buildings in Singapore MPC with weather forecasting capabilities, whereas
13
have achieved more than 65% improvement in energy Ma et al. achieved a 19% efficiency improvement
efficiency compared to 2005 levels. The BCA further aims using a hierarchical MPC system for a chiller plant. In
14
to raise the energy efficiency improvement of Singapore’s experimental studies, Pang et al. demonstrated that MPC
buildings to 80% relative to 2005 levels by 2030 through its reduced chilled water consumption by 42% in a radiant
Green Building Innovation Cluster program. 3 slab system compared to heuristic methods. Despite
these promising results, two primary obstacles hinder the
The conventional approach in current building widespread adoption of MPC: the need for highly accurate
automation and control systems is predominantly predictive models and the significant computational cost
“reactive,” generating control signals based on deviations associated with solving complex optimization problems
of previously measured information from a control across multiple zones. To address these issues, Yang et al.
15
setpoint (thermostat-based). Consequently, due to the proposed a two-level distributed computation scheme that
different thermal characteristics of buildings and the optimizes thermal comfort and internal air quality across
non-linear operation of their ACMV systems, reactive multiple zones using MPC. In addition, an event-triggered
control strategies are unable to achieve optimal efficiency optimization mechanism has been shown to reduce the
and occupant comfort. These strategies only respond to computational load by 12–22% while achieving over 9%
transient disturbances – changes in occupancy, internal energy savings, further highlighting MPC’s potential in
16
heat load profiles, and external weather conditions – as real-world applications.
well as to building dynamics after the disturbances occur.
This limitation highlights the necessity for predictive Alongside MPC, recent advancements in machine
control strategies that can anticipate and mitigate such learning (ML) have introduced powerful tools for
disturbances before they adversely impact building predicting building energy performance. A review by Fathi
17
performance. et al. examined 14 ML methods relevant to this domain.
Among these, support vector machines (SVM) are
Studies on predictive control solutions for optimizing particularly effective for regression tasks involving non-
ACMV operation in recent decades can broadly be linear relationships, especially when using the radial basis
categorized based on model architecture into (i) model function kernel, which captures non-linear interactions
predictive control (MPC) and (ii) reinforced learning more effectively compared to sigmoid and polynomial
(RL). MPC uses prediction models to forecast future kernels. 18,19 Random forest (RF), an ensemble learning
4
system behavior and proactively adjust control actions, method based on decision trees, is adept at managing non-
making it particularly well-suited for complex systems linear interactions and offers improved prediction stability,
with non-linear dynamics. Within the MPC framework, albeit at the expense of increased computational time.
20
three primary techniques are used to develop prediction In the realm of artificial neural networks (ANN), the
models: physics-based (white box), data-driven (black non-linear autoregressive network with exogenous inputs
box), and hybrid (gray box) methods. Physics-based (NARX) is well-suited for dynamic systems, leveraging
5
models rely on fundamental principles such as mass both historical and external inputs to forecast future
and energy conservation and can be implemented using states. In addition, long short-term memory (LSTM)
21
software tools such as EnergyPlus, TRNSYS, eQuest, networks, a type of recurrent neural network known
8
7
6
and Modelica. However, these models require significant for capturing long-term dependencies, are particularly
9
expertise, resulting in higher implementation costs, effective for modeling the time-series dynamics of building
particularly in complex real-world buildings. In contrast, energy consumption. Integrating these ML-based
22
data-driven models apply regression techniques to large forecasting methods with MPC can further enhance the
Volume 2 Issue 1 (2025) 40 doi: 10.36922/ijamd.8161