History matching predictive control for HVAC chiller sequencing and comfort-aware cooling optimization

Buildings in Hong Kong account for nearly 90% of the city's total electricity consumption, with HVAC systems representing the largest share and incurring annual costs exceeding HK$12.3 billion. A primary source of inefficiency is improper chiller sequencing during transitional weather, where static control strategies fail to respond promptly to temperature drops, leading to systemic overcooling and significant energy waste. To address this, this study proposes a predictive control framework for dynamic chiller sequencing. The methodology integrates a Scalable Dynamically Engineered Multi-Modal Feature Learning (Scalable DEMMFL) model for high-precision cooling load prediction with a History-Matching Cooling Classification (HMCC) method to categorize thermal states into under-, balanced-, or over-cooling. The framework was validated using real-world operational data from two distinct commercial buildings in Hong Kong during transitional seasons characterized by frequent temperature fluctuations. Experimental results indicate that the Scalable DEMMFL model provides superior predictive performance, achieving a 20% reduction in Root Mean Squared Error (RMSE) compared to established baseline models. Furthermore, the implementation of the proposed framework effectively mitigated overcooling, resulting in a cooling load reduction of approximately 35,000 kWh for Building A and 10,000 kWh for Building B across identified overcooling days. These findings demonstrate that the framework facilitates dynamic, temperature-responsive control, effectively reducing energy consumption while maintaining indoor thermal comfort and advancing smart building energy management.