Resilient coordination of energy storage and renewable generation in residential systems using deep reinforcement learning considering extreme weather conditions

The increasing frequency of extreme weather events presents critical challenges to the resilience of energy distribution systems and the operation of energy management systems (EMS). Traditional model-based optimization approaches rely heavily on accurate forecasts and detailed system models, which can be difficult to obtain under high uncertainty. To address this limitation, this study proposes a data-driven EMS framework based on deep reinforcement learning (DRL), which learns robust coordination strategies for residential energy resources directly from stochastic environments. The framework is designed to optimize energy coordination in residential buildings during both normal and extreme operating conditions. The proposed framework combines renewable generation, mobile and stationary energy storage, and employs a sequential decision-making model to adapt its actions in real time based on system dynamics. The DRL agent is trained using event-adversarial scenarios to enhance robustness and resilience in unpredictable environments. Using real household data from Queensland, Australia, we validate the agent's ability to maintain load support, minimize energy not supplied, and improve renewable generation self-consumption. Compared to deterministic optimization methods, the DRL-based approach achieves comparable performance with faster response, requiring minimal user intervention. This makes it highly suitable for real-time EMS deployment in weather-vulnerable urban areas. The proposed method offers a scalable and intelligent solution to enhance the resilience, sustainability, and energy efficiency of residential systems in cities facing uncertain and extreme climatic conditions.