LightWind: A Lightweight Framework for Efficient Wind Power Forecasting

Wind power forecasting is crucial for the efficient operation and stability of power grids. However, in practical applications, forecasting models encounter limitations in computational costs when deployed on wind farms located in high-latitude regions or offshore areas. Additionally, these models suffer from inadequate dynamic modeling of turbine-to-turbine interactions within wind farms, which affects their accuracy and scalability in large-scale deployment scenarios. To address these challenges, this paper introduces LightWind, a lightweight wind power forecasting framework that considers the interactions between different turbines within a wind farm, designed for efficient deployment on resource-constrained devices while significantly improving forecasting accuracy. The framework utilizes a cosine similarity algorithm to quantify the influence between turbines and adopts a lightweight spatiotemporal operator structure, integrating a Shuffled Group Temporal Convolution Network (SG-TCN) and a Space-Fusion Transformer (SF-Former). These components efficiently extract spatiotemporal features of the wind farm while maintaining low computational overhead. Furthermore, LightWind incorporates a feature compression mechanism that effectively reduces computational redundancy and accelerates real-time prediction. Experimental results on benchmark datasets for both single-step and multi-step forecasting demonstrate that LightWind achieves superior forecasting accuracy compared to state-of-the-art methods, while significantly reducing computational and storage costs.