This study develops an explainable reinforcement learning (RL) control framework for hybrid ventilation in Mediterranean office buildings to enhance thermal comfort, energy efficiency, and long-term climate resilience. A working environment was created Using EnergyPlus to represent an office test cell equipped with natural ventilation and air conditioning. The RL controller, based on Proximal Policy Optimization (PPO), was trained exclusively on present-day Typical Meteorological Year (TMY) data from Beirut and subsequently evaluated, without retraining, under future 2050 and 2080 climate projections (SSP1-2.6 and SSP5-8.5) generated using the Belcher morphing technique, in order to quantify robustness under projected climate stressors. Results showed that the RL control achieved consistent, though moderate, annual HVAC energy reductions (6–9%), and a reduction in indoor overheating degree (IOD) by about 35.66% compared to rule-based control, while maintaining comfort and increasing natural ventilation hours. The Climate Change Overheating Resistivity (CCOR) improved by 24.32%, demonstrating the controller’s resilience under warming conditions. Explainability was achieved through Kernel SHAP, which revealed physically coherent feature influences consistent with thermal comfort logic. The findings confirmed that physics-informed RL can autonomously learn and sustain effective ventilation control, remaining transparent, reliable, and robust under future climates. This framework establishes a foundation for adaptive and interpretable RL-based hybrid ventilation control, enabling long-lived office buildings in Mediterranean climates to reduce cooling energy demand and mitigate overheating risks under future climate change.
Krayem et al. (Tue,) studied this question.