Modality alignment-driven large language model for wind farm power forecasting

Accurate forecasting of wind farm power output plays a vital role in grid dispatching, optimizing transactions in electricity markets, and enhancing the overall operational security of power systems. In recent years, large language models (LLMs) have been introduced into wind power forecasting due to their strong reasoning capabilities and rich domain knowledge. However, applying LLMs in this domain remains challenging, primarily because they require substantial amounts of data and impose stringent deployment requirements. In practice, the harsh environments in which wind farms operate make it difficult to collect sufficient power data to meet the demands of LLM training. Moreover, real-world deployment often falls short of the computational and infrastructural requirements these models entail. As a result, LLM-based forecasting methods have struggled to meet the expectations for accuracy and practicality in real scenarios. To address these challenges, a novel modality alignment-driven large language model (MALLM) for wind farm power forecasting is proposed. The proposed model leverages a modality alignment method to resolve the mismatch between the modality of power data and the input expectations of LLMs. It eliminates the need for retraining or fine-tuning the language model, significantly reducing both deployment costs and data requirements. Furthermore, the modality alignment method incorporates temporal information embedding, enabling the integration of aligned time-related features with power data. This helps preserve key temporal patterns that are often lost during standard modality transformation processes. Comparative and ablation experiments conducted on supervisory control and data acquisition (SCADA) datasets from two wind farms located in distinct geographical regions demonstrate the effectiveness and robustness of the proposed method.