A trustworthy hybrid human-LLM assisted safe reinforcement learning method for distribution network operation

This is the authors' abstract. We don't add key points for this paper.

Distribution network is transitioning from model-based to data-driven operations amid increasing integration of renewable energy sources, posing new challenges for operational safety and cost-effectiveness. Reinforcement learning (RL), characterized by its strong adaptability and fast response, shows significant promise in this context. However, RL suffers from poor generalization against various uncertainties and heavy labor cost due to the need for extensive model or neural network modifications and hyperparameter tuning. More importantly, once deployed, the performance of RL algorithms is highly dependent on domain expert knowledge for specific power system operation or control tasks, often relying on unrealistic assumptions of perfect knowledge acquisition. Large language models (LLMs) have achieved notable success in numerous areas and present a potential approach to replace domain expert guidance, though it raises new challenges such as hallucination and limited interpretability. To address these challenges, this paper proposes a hybrid human–LLM assisted RL framework that balances human supervision with LLM assistance to improve training generalization, efficiency and stability amid high-variance uncertainties. Furthermore, a coordination-enhanced safe multi-agent RL algorithm is introduced to improve compliance with global distribution network operating constraints under decentralized control schemes among interconnected energy assets. Case studies conducted on standard distribution test systems with a domain knowledge quantification method show that the proposed method reduces voltage violations, achieves performance within approximately 5% of the theoretical optimum under unseen scenarios, and a pilot evaluation with real human participants suggests a substantial reduction in labor costs compared to purely human-driven approaches. Overall, the proposed hybrid human–LLM assisted safe RL (HH-LLM-SRL) method offers a trustworthy solution for intelligent and safe operation of distribution networks and introduces practical labor cost as a new evaluation dimension for LLM-assisted energy system decision-making.