AI-driven climate-resilient HVAC control: A scalable framework for grid-Interactive buildings

Residential cooling systems in extreme temperature regions often operate at maximum capacity for extended periods and creating significant challenges for conventional control strategies. Conventional control strategies fail to manage the extreme diurnal temperature variations (>19°C) and rapid environmental transitions characteristic of these regions, resulting in excessive energy consumption and compromised thermal comfort. This study presents a field-validated framework integrating Digital Twin modeling, Graph Neural Networks, and Deep Reinforcement Learning for adaptive building energy management under extreme climatic conditions. The methodology was evaluated through a 14-month field study across five residential buildings in Hail, Saudi Arabia (BWh climate classification), experiencing summer temperatures exceeding 45°C. The DT model achieves high-fidelity thermal predictions (temperature RMSE: 0.62°C; energy RMSE: 0.35 kWh), enabling realistic virtual commissioning of control strategies. The GNN component models inter-zone thermal dynamics with 0.28°C RMSE for 1-h forecasts, capturing spatial dependencies conventional approaches neglect. The DRL controller optimizes HVAC operations through continuous learning, adapting to occupancy patterns and environmental conditions. Results demonstrate 35.3% energy reduction compared to conventional thermostatic control (95% CI: 32.1–38.5%), while maintaining thermal comfort within acceptable ranges for 92.0% of occupied hours versus 68.2% for baseline systems. Performance remains robust during extreme events, achieving 42.5% savings during dust storms through anticipatory control strategies. Economic analysis indicates a 26.3-month payback period with 143% five-year return on investment. The framework demonstrates scalability potential, with coordinated control across 100 buildings increasing savings to 38.0% through load balancing and demand coordination.