Field-validated climate-resilient HVAC control for mixed-use buildings in a hot-arid region

Buildings in desert cities are difficult to operate efficiently because heat waves and airborne dust increase cooling demand, alter short-term thermal behavior, and reduce HVAC efficiency. This study develops and field-validates an adaptive control framework for mixed residential and commercial buildings in a hot-arid climate. The framework integrates a continuously calibrated digital twin, a heterogeneous spatiotemporal graph neural network for short-horizon thermal and load prediction, and decentralized multi-agent deep reinforcement learning. The control architecture uses Proximal Policy Optimization and a shared graph embedding to update HVAC setpoints. Its comfort constraints align with ASHRAE Standard 55 and the high-performance building objectives of ASHRAE Standard 189.1. The framework was evaluated for 18 months in eight occupied buildings in Hail, Saudi Arabia: six residential units and two small commercial facilities. The evaluation included periods with outdoor temperatures above 45 °C and 48 recorded dust-storm events. Compared with conventional thermostat operation, the controller reduced electricity use by 37.8% and maintained the ASHRAE PMV comfort band during 93.5% of occupied hours. Energy savings were 36.5% in residential buildings and 41.8% in commercial buildings. During dust events, anticipatory control increased these savings to 44.2% and 46.8%, respectively. The framework also outperformed model predictive control and a local DRL baseline without graph-based coordination. Transfer learning reduced commercial training time by 38% and computational demand by 56%. Separate community-scale simulations based on the calibrated digital twin projected 41.2% energy savings and a 43.8% peak-load reduction for 150 buildings; these values were not obtained from an additional field deployment. The field results show that climate-specific AI control can improve energy efficiency and operational resilience in cooling-dominated buildings under severe desert conditions. Validation in larger and more diverse building portfolios is still required.