Large Language Model Based Data Augmentation for Peak Electricity Price Forecasting and Battery Energy Storage Arbitrage

Battery Energy Storage Systems (BESS) play a vital role in enhancing grid flexibility, enabling renewable integration, and supporting peak shaving and frequency regulation. To fully realize their economic potential, BESS operators often participate in electricity market arbitrage - charging when prices are low and discharging during peak price periods. However, effective arbitrage strategies critically depend on accurate forecasting of electricity prices, particularly under extreme market conditions, which are often driven by sudden load surges, grid failures, or severe weather events. In this paper, we propose a novel framework that leverages a Large Language Model (LLM) for data augmentation to address the scarcity of extreme price scenarios in historical data. The LLM agent, guided by structured prompts that embed domain knowledge and physical constraints, generates realistic synthetic samples of rare peak-price events. These augmented datasets improve the robustness of a Bayesian electricity price forecasting model based on Monte Carlo dropout, which provides not only point estimates but also predictive confidence intervals. Finally, a scenario-based stochastic optimization model is developed to guide BESS arbitrage decisions using the probabilistic price forecasts. Simulation results show that the proposed framework significantly enhances the predictive accuracy and economic efficiency of storage arbitrage under uncertainty. © 2025 IEEE.