Weather-Aware Reinforcement Learning for Power Grid Topology Control Under Extreme Events

Abstract

Extreme weather and high renewable penetration increase the vulnerability of power system operations to nonstationary, heavy-tailed disturbances. Most deep reinforcement learning (DRL) controllers for topology control are trained under nominal conditions and treat weather as a latent disturbance, leading to sharp performance degradation or even worse-than-Do-Nothing behaviour in extreme regimes. This paper proposes a weather-aware soft actor–critic (WA-SAC) framework that makes weather an explicit, structured input to corrective topology control in RES-rich grids. The framework couples: (i) an LLM-driven extreme-weather data augmentation pipeline that generates scenario-oriented, physically consistent renewable and load trajectories via lifecycle modulation and two-pass load balancing; (ii) an 8-dimensional hybrid weather encoding (6D categorical type, intensity, and trend) and a weather-conditioned graph-attention actor–critic that learns adaptive topology actions; and (iii) a deployment-oriented dual-path weather inference scheme supporting the workflow of training with oracle weather codes and deploying with inferred weather codes. Experiments on Grid2Op WCCI 2022 (IEEE 118-bus) with eight extreme-weather categories show that LLM-generated data is 11.7% more challenging than scaling-based alternatives on Do-Nothing survival; weather encoding improves extreme-scenario performance by 18.6% over the no-encoding baseline with scenario-dependent behaviour in mild regimes; and LLM-inferred weather codes achieve an 88% functional match rate to oracle performance under a 10% degradation tolerance, validating the proposed workflow for deployment.