A Robust Safe Reinforcement Learning Approach for Power Grid Resilience Enhancement against Typhoons via DER Flexibility Aggregation

Climate change has led to an increase in both the frequency and severity of weather-related power outages globally. The inherent stochasticity of extreme weather phenomena has substantially disrupted power grid operations, intensifying vulnerabilities arising from computational complexity, uncertainty, and reduced system inertia, particularly in networks with high penetration of distributed energy resources (DERs). To mitigate these challenges, this study introduces a resilience enhancement framework for distribution systems based on robust safe reinforcement learning (RSRL), which exploits the aggregated flexibility of DERs under frequency security constraints. Initially, acknowledging the spatiotemporal effects of typhoons on distribution networks, a flexibility aggregation model is developed employing linear approximation and adaptive robust optimization (ARO) to reliably define the safe dispatchable range of DERs under both normal and extreme weather scenarios. Subsequently, a novel grid resilience enhancement model is formulated, integrating analytically derived and linearized frequency security constraints that comprehensively incorporate frequency security margins to effectively alleviate typhooninduced impacts. Moreover, to address the dynamic and timevarying solution space encountered during disaster events, a robust safe reinforcement learning methodology is proposed, facilitating efficient resolution of models characterized by complex constraints and uncertainties related to source-load coupling. Validation through simulations on the modified IEEE 30-bus system demonstrates the efficacy of the proposed approach in substantially improving system resilience and stability, while effectively managing challenges associated with extreme weather conditions and frequency stability.