CMLLM: A novel cross-modal large language model for wind power forecasting

Abstract

Accurate short-term wind power forecasting is crucial for ensuring grid stability and optimizing the operation of wind farm-energy storage systems. However, the inherent randomness and high variability of wind energy present significant challenges to wind power forecasting. To leverage the powerful reasoning capabilities and high-level knowledge of large language models for accurately extracting features from non-stationary wind power data, a cross-modal large language model (CMLLM) is proposed for wind power forecasting. This model employs data cross-modal and pre-trained large language models, enabling efficient compatibility with various large language models and adaptability to data with different characteristics. In CMLLM, data is comprehensively processed through the adoption of a cross-modal transfer learning method. Data is converted into text modality, thereby eliminating the need for re-training or fine-tuning large language models, while effectively mitigating the decline in forecasting accuracy due to modal mismatches. A prior knowledge prompt prefix module is designed in the forecasting process to enhance feature extraction and activate the reasoning capabilities of the large language model for wind power forecasting tasks. Extensive experiments are carried out on datasets from three wind farms in China to evaluate the performance of CMLLM. The results validate the model's effectiveness, accuracy, and robustness across diverse datasets and a range of large language models.