Comfort-Gated Supervisory Control of an Office HVAC Zone in a Warm–Humid Climate: Gradient-Boosted Surrogates with Grid Search and Particle Swarm Optimization

Abstract

HVAC plant dominates commercial building electricity demand, and in warm–humid climates the latent component is large enough to overturn control strategies designed around temperature alone. This paper presents a supervisory control framework in which three gradient-boosted-tree surrogates — zone temperature, relative humidity and hourly energy — are trained on one year of measured building-management-system data augmented by a calibrated psychrometric plant, and then serve as both the optimisation model and the evaluation plant for four supervisory controllers. Two perform exhaustive per-hour grid search over a discrete setpoint–dehumidifier action space; two tune a four-parameter operating schedule by particle swarm optimisation (PSO). Each appears in a night-setback and a no-setback variant under identical hardware-realism constraints: integer setpoints, a two-hour minimum compressor hold, occupancy-based operation and pre-cooling look-ahead. Selection is not made on energy — a three-part comfort gate (occupied-hour temperature compliance, humidity compliance, Fanger-based satisfaction) must be satisfied before a controller is eligible, and only the highest-saving eligible controller is locked and serialised.

Over a full 2025 evaluation year for an office zone, the surrogates reproduced measured dynamics at R² = 0.998, 0.994 and 0.999. The grid-search controllers achieved the largest nominal savings, 19.0% and 19.8%, but both failed the humidity gate at 79.2% and 83.4% compliance and were rejected. The PSO schedule with night setback was locked at 6.0% annual saving 100% temperature compliance, 96.6% humidity compliance, 96.6% of occupied hours inside the ASHRAE comfort box and no occupied hour above PPD 20; cumulative occupied discomfort fell by ~58%.