Large Language Model Compatibility With Reinforcement Learning for Networked Microgrids Considering Device and System-Level Missing Measurements

The deep reinforcement learning (DRL) approach with its end-to-end and data-driven features enhances the operational strategies for networked microgrids (NMGs). Well-trained DRL agents can make optimal decisions for system operations by observing NMG online measurements. However, missing measurements pose an unpredictable challenge for the secure operation of DRL-based NMGs. This paper proposes a large language model (LLM)-based agent designed for compatibility with multi-agent DRL online decision-making for networked microgrids considering missing measurements. The novel design rules for the LLM-based agent are proposed to unleash the potential of LLM in imputing missing NMG measurements using only few-shot learning. Subsequently, the designed LLM-based agent is embedded into multi-agent DRL in a compatible manner for NMG online operations considering missing online measurements. Experimental results indicate that the proposed method exhibits robustness and trustworthiness in device and system-level measurement imputations while holding the merits of DRL. The proposed method reduces NMG's total operation costs by up to 23.33%, achieving a balance between security and optimality under missing measurements.