PEFT-based large language model for wind power forecasting and risk-tunable energy scheduling in microgrids

The increasing penetration of uncertain wind power in microgrids challenges scheduling, security, and reliability. To enhance forecast performance, this paper proposes a new method utilizing a large language model (LLM) with parameter-efficient fine-tuning (PEFT) technique, specifically Low-Rank Adaptation (LoRA) for wind power forecasting. Leveraging LLMs’ superior pattern recognition and contextual understanding capabilities, the historical wind power data, meteorological inputs, and relevant context are used in text-based prompts via a supervised instruction fine-tuning strategy. This enables the model to capture complex dependencies and predict future output. Building on the LLM forecasts, a data-driven, risk-tunable uncertainty set based on the Wasserstein distance is developed to characterize forecast errors. This set incorporates the operator’s risk preference. It feeds a robust optimization framework for microgrid scheduling, enabling optimization at varying risk levels. The prompt-optimized LLM forecasting model achieves a 37.99% reduction in mean squared error versus traditional LSTM on the target dataset. Furthermore, the derived risk-tunable uncertainty set enables a robust optimization framework, validated on a modified IEEE 39-bus system, that reduces operational cost by about 5.53% compared to the traditional robust method and improves computational efficiency by about 94.96% versus stochastic optimization, maintaining an effective conservatism-efficiency balance.