PowerMistral: A data-efficient wind power forecasting framework leveraging pre-trained large language models

As energy systems decarbonize worldwide, wind power has become a pillar of sustainable development with rapidly expanding capacity; however, many new wind farms lack historical operational data, which can impair the accuracy and reliability of traditional forecasting models. To address this challenge, we present PowerMistral, a novel framework for few-shot wind power forecasting. The framework is based on Mistral-7B, a small pre-trained large language model that can efficiently complete prediction tasks without requiring a lot of retraining. Currently, PowerMistral uses a two-step fine-tuning process. While targeted fine-tuning increases prediction accuracy in the second stage, time-series feature extraction captures the temporal characteristics of wind power in the first phase. To improve the capacity for pattern recognition and analysis, a novel Gated Multi-scale Convolutional Neural Network (GMCNN) is proposed to capture the unique short-term fluctuations and long-term trend patterns inherent in wind power sequences, while also incorporating textual prompts related to these trends. Additionally, advanced time encoding techniques and domain-specific numerical features are integrated with the aforementioned patterns. Improved performance, especially in small-sample situations, was demonstrated when the framework's effectiveness was evaluated on six operational wind farms in Tianjin, China. Notably, the largest observed reductions in MAE and RMSE reached approximately 75.53% and 75.43%, representing the maximum performance gains achieved among the comparative models.