Dynamic Mode Decomposition-Enhanced Deep Reinforcement Learning for Resilient Multi-Zone HVAC Control

Modern Heating, Ventilation, and Air Conditioning (HVAC) systems require robustness against persistent disturbances—including sensor drifts, actuator degradations, and unforeseen faults—to sustain indoor comfort and energy efficiency. This paper proposes a resilient hybrid framework integrating Dynamic Mode Decomposition (DMD) with Deep Reinforcement Learning (DRL) for adaptive multi-zone HVAC control. A multivariate DMD module continuously learns low-dimensional linear dynamics from streaming sensor data, enabling real-time self-adaptation to non-catastrophic faults. A Deep Q-Network (DQN) agent leverages this model to maintain near-optimal control policies, preserving performance without fallback activation during everyday degradations. For severe faults (detected via DMD residuals), a fail-safe PID controller ensures stability. Simulation results on a three-zone building model confirm the framework’s resilience: it significantly reduces comfort violations compared to pure DRL baselines while maintaining energy efficiency close to optimal, demonstrating robust adaptation under fault conditions.