Enhancing Grid Flood Resilience with CFD Fragility Modelling and Risk-Constrained DRL

Flood-induced failures in power distribution networks compromise grid resilience, while existing resilient control strategies often rely on static risk models and lack integration with physical flood dynamics. This paper presents a hybrid framework that combines Computational Fluid Dynamics (CFD) with Deep Reinforcement Learning (DRL) to enable dynamic risk-aware control under evolving flood conditions. High-resolution flood scenarios are simulated using ANSYS Fluent, where stochastic hydrodynamic parameters are sampled from calibrated Weibull distributions. These parameters are embedded into a Maskable Proximal Policy Optimisation (PPO) agent trained with Conditional Value-at-Risk (CVaR) constraints, enabling the learning of risk-constrained control policies without requiring real-time CFD computations. This architecture preserves physical fidelity while ensuring dynamic feasibility. When evaluated on the extended IEEE 123-bus feeder, the proposed CFD–DRL controller achieves a 49% reduction in energy not supplied (ENS) and improves CVaR compliance from 85% to 98%, compared to the conventional MPC baseline. The results demonstrate that integrating high-fidelity numerical modelling with scalable learning-based control significantly enhances the adaptability and robustness of flood-resilient grid operations.