Hybrid ML-LTI control of AHU supply air temperature with comfort-energy metrics

Automated supply air temperature (ASAT) is a critical parameter in modern commercial buildings, directly influencing indoor climate. Recent advances in machine learning (ML) have shown strong performance in building applications, yet their limitations and the trade-offs between performance and implementation cost remain case-dependent. This paper investigates the modeling and control of supply air temperature in air handling unit (AHU) systems using a hybrid approach that combines ML with classical control techniques. The proposed method integrates classical linear time-invariant (LTI) system analysis with a practical, data-driven implementation — specifically, a Linear Regression model with Huber loss, augmented by a structured Hankel-matrix feature set — to derive an optimal control curve tailored to a specific AHU. A comprehensive control framework comprising three activation stages — ML-only, rule-based, and mixed — is developed and detailed. The method is validated using real data from operational AHUs in a commercial office building, with automatic data mining and pre-clustering used to identify historical periods of optimal comfort. Performance is compared against a rule-based control strategy under similar weather conditions, showing significant improvements across several key performance indicators, including a comfort-corridor success rate and an energy metric defined as the work needed to maintain that corridor. While the ML-only stage demonstrates low computational overhead, it requires more post-exchanger heating energy than the rule-based or mixed strategies, guiding future optimization efforts.