Page 55 - IJAMD-2-1

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

fans remained approximately the same for both baseline occupant comfort in a multizone commercial building. The
BMS and MPC. During MPC duration, it was found that implemented system achieved energy savings exceeding
approximately 32% of the total cooling energy consumption 42% compared to conventional thermostat-based control
was attributable to electric fan power. If FCU fans equipped while simultaneously improving thermal comfort and
with variable-speed drives were installed instead of constant- stability. The novelty of our approach lies in its real-world,
speed fans, MPC would have the opportunity to further large-scale deployment in an 11-zone building, integration
optimize fan power consumption. For example, MPC of ML-based weather forecasting for proactive control
could be configured to reduce FCU fan speed by over 70% adjustments, and the use of a NARX neural network for
compared to constant-speed operation when no occupants accurate PMV prediction.
are present in the test rooms. By adopting such control logic While initial implementation costs and model
for optimizing fan operation, a conservative reduction of development efforts present challenges for widespread
10 – 15% in fan power consumption was estimated. This adoption, future research endeavors aim to mitigate these
estimation excludes the reduction in cooling energy produced limitations. Such efforts include developing comfort models
by chilled water (which can be measured by the BTU meter) that rely exclusively on existing building sensors, exploring
during the heat exchange process with the reduced airflow cost-effective sensor technologies, and examining the
in the FCU. Such a detailed impact of reduced fan power on integration of additional building systems, such as lighting
cooling energy consumption will be investigated in future
studies. Excluding the potential effect of airflow reduction on and shading, for more holistic optimization strategies. The
chilled water cooling energy, an additional 3 – 5% reduction present study contributes to advancing intelligent building
in total cooling energy consumption was estimated. automation systems, with significant potential to enhance
energy efficiency and sustainability in the built environment.
A key advantage of our MPC implementation is its
modular and adaptable framework, involving: (i) collecting Acknowledgments
building data; (ii) developing a control-oriented model Technical support from Dr. Wai Soong Loh from JTC for
using ML tools; (iii) designing the optimization algorithm; this project is much appreciated. We also appreciate the
and (iv) integrating the MPC with the BMS. While this patience, understanding, and logistical support of the
framework is generally applicable to buildings in various administrative team from Civil Service College, Singapore,
climates and with diverse operational profiles, building- for this project.
specific retraining is necessary to accurately capture unique
thermal dynamics. In environments with significant seasonal Funding
or regional variations, model parameters (e.g., lag length for
NARX or input depth for LSTM) may require adjustment, This research is jointly supported by Jurong Town
posing challenges such as increased computational cost and Corporation (JTC) of Singapore (NTU REF 2019-0607)
potential overfitting. Future studies will further validate the and Smart Nation and Digital Government Office, SNDGO
framework across different climatic zones. of Singapore (NRF2016IDM-TRANS001-031).
A key challenge in MPC deployment is balancing model Conflict of interest
complexity against computational cost. In our current
implementation, the NARX model provided superior The authors declare they have no competing interests.
accuracy and efficiency (Table 4). For more complex Author contributions
systems, LSTM architectures – with increased input depth
and larger hidden layers – could offer enhanced prediction Conceptualization: Pradeep Shakya, Man Pun Wan
accuracy, albeit with higher computational demands. In such Data Curation: Shiva Sreenivasan
cases, employing artificial intelligence/ML-optimized edge Formal analysis: Pradeep Shakya, Man Pun Wan
hardware and multithreaded computing can help manage Funding acquisition: Shiyu Yang, Man Pun Wan
computational loads. In our previous work, we introduced an Investigation: Shiva Sreenivasan, Pradeep Shakya
11
approach using instantaneous linearization within ML-based Methodology: Pradeep Shakya
MPC to further address these challenges. Although a detailed Resources: Baskaran Krishnamoorthy, Shiyu Yang
exploration of this topic is beyond the scope of the present Writing – original draft: Pradeep Shakya, Shiva Sreenivasan
paper, it represents a promising direction for future research. Writing – review & editing: Man Pun Wan, Pradeep Shakya,
Shiva Sreenivasan
6. Conclusion
This study demonstrated the effectiveness of an ML-based Ethics approval and consent to participate
MPC system in optimizing energy consumption and Not applicable.

Volume 2 Issue 1 (2025) 49 doi: 10.36922/ijamd.8161