Machine learning-based model predictive control for multizone building automation: A case study

In Singapore’s hot and humid climate, air-conditioning and mechanical ventilation (ACMV) systems account for over 60% of commercial building energy consumption, driving efforts to enhance energy efficiency through predictive control strategies such as model predictive control (MPC) to overcome the limitations of conventional reactive building automation systems. This paper presents a multizone MPC system designed to optimize energy consumption and thermal comfort in a commercial building’s ACMV system in Singapore. The system was implemented in a multi-use test building with real occupancy and a deployment area of approximately 850 m2, partitioned into six learning zones, two office spaces, and three open spaces. The ACMV system serving the deployment area consisted of two primary air-handling units and 16 fan coil units, where chilled water was supplied to the cooling coils, and conditioned air was distributed through motorized diffusers. To facilitate predictive control, data-driven thermal prediction models were developed for each zone using a non-linear autoregressive exogenous network with exogenous inputs trained on historical data and disturbances. Thermal comfort optimization was guided by the predictive mean vote, which was targeted at 0, representing thermal neutrality (as per ASHRAE 55 standards), and constrained within a range of −0.5 – 0.5. Performance comparisons demonstrated that the MPC system achieved over 42% energy savings compared to the original thermostat-based control while enhancing thermal comfort. Despite its advantageous control performances, challenges for large-scale deployment remain, including implementation costs, scalability, and model accuracy. Future work can address these challenges by developing comfort models that leverage existing building sensors.