Enhancing power grid resilience through weather-aware security constraints: A deep reinforcement learning approach with hybrid CNN-GRU architecture

Extreme weather events increasingly challenge the operational resilience of distribution systems by introducing dynamic and uncertain security limits (SLs), alongside data sparsity. Traditional model-based approaches often rely on static assumptions and require complete system modeling, making them difficult to adapt to rapidly evolving weather-induced constraints. To address these limitations, this paper proposes a model-free resilience enhancement framework based on deep reinforcement learning (DRL), integrating real-time weather-aware SL identification and adaptive dispatch. First, an ensemble Bagging-XGBoost model is developed to classify weather severity levels and determine whether static or dynamic SLs should be applied, enabling scenario-adaptive SL switching. Second, a hybrid convolutional neural network–gated recurrent unit (CNN-GRU) model, enhanced by transfer learning, is designed to accurately estimate dynamic SLs under varying weather conditions. The CNN captures spatial meteorological patterns, while the GRU models temporal evolution; transfer learning improves generalization under limited training data. Third, the dispatch problem is formulated as a constrained Markov decision process (CMDP), and solved using a primal–dual deep deterministic policy gradient (PD-DDPG) algorithm that explicitly incorporates SL constraints into the policy learning process. An attention-based meteorological data reconstruction model is further integrated to enhance the quality of input data and training efficiency. Case studies on the improved IEEE-123 test feeder demonstrate that the proposed method reduces average load loss by 23.30 % and 12.10 % compared to CNN-only and GRU-only baselines, respectively. Moreover, it achieves an 88.77 % improvement in computational efficiency over conventional model-based resilience strategies, highlighting its robustness and applicability under limited data and high-impact weather conditions.