Causality-Aware LLM-Enhanced Graph Representation Learning for Adaptive Power System Control

High renewable penetration and reduced system inertia introduce significant challenges for transient stability assessment and control. This article proposes a causality-aware, large language model–enhanced distribution-preserving graph representation learning framework (LLM-DP-GRL) for fast and accurate stability prediction and decision-making. The DP-GRL model captures both structural and distributional properties of network states, whereas large language models provide physics-informed priors that improve data efficiency and generalization under multicontingency and out-of-distribution scenarios. A causal intervention module further quantifies bus-level influence on stability margins, offering interpretable insights consistent with system dynamics. The learned surrogate model is integrated into a cooperative preventive–emergency control strategy, enabling real-time stability margin evaluation and optimization. Tests on the IEEE 39-bus and 118-bus systems show that LLM-DP-GRL achieves higher accuracy, faster convergence, and improved robustness compared with conventional machine learning, LSTM, and GNN-based methods. The proposed approach reduces online control computation from over 35 min (TDS-based) to 39 s while maintaining inference latency below 30 ms. These results demonstrate that combining graph learning, LLM-guided priors, and causal analysis provides an effective and scalable solution for stability assessment and emergency control in low-inertia, high-renewable power systems.

© 2026 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission.