🤖 AI Summary
To address the challenge of jointly optimizing grid security and traffic efficiency in real-time EV charging recommendations under dynamic traffic–power system coupling, this paper proposes a prediction-guided safe reinforcement learning framework. The method integrates graph neural networks (GNNs) to model the coupled topology, temporal convolutional networks (TCNs) to forecast supply–demand evolution, and a proximal policy optimization (PPO) algorithm augmented with an explicit safety barrier to ensure online decision-making safety, robustness, and formal verifiability. Evaluated on a city-scale co-simulation platform, the framework achieves a 23.6% improvement in charging scheduling success rate, reduces voltage limit violations by 91.4%, and maintains an average decision latency below 85 ms—marking the first demonstration of millisecond-level, safety-guaranteed adaptive charging recommendations.