An In-Context LLM for PV-BESS Operations: Adaptive Day-Ahead Strategy Recommendation for Economic Optimization

초록

This study presents a strategy recommendation methodology for the optimal operation of photovoltaic-battery energy storage systems (PV-BESS) by leveraging large language models (LLMs) through in-context learning (ICL) and prompt engineering. Unlike traditional rule-based approaches, the proposed method effectively integrates electricity tariffs, solar PV generation, load, battery degradation, and policy constraints to generate context-specific operational strategies automatically. A 10-year simulation of a South Korean campus building indicated that the LLM-based approach reduced cumulative operating costs by 5-12% and lowered life cycle cost (LCC), including initial investment, replacement, and residual value, by 3-5%, demonstrating long-term economic viability. GPT-2 achieved reduced battery degradation through conservative operation, whereas LLaMA 3-8B improved PV utilization and renewable energy certificate (REC) revenue via diversified strategy selection. The findings confirm that LLM-based methods can be developed into intelligent operational systems that simultaneously optimize economic performance, efficiency, and sustainability, with potential applicability to distributed energy resource management and smart grid operations. © 2013 IEEE.

키워드