WindPower-SAFusion: A Sparse-Attention and Multi-Scale Fusion Model for Wind Power Forecasting

With the increasing penetration of wind power, the uncertainty of wind power generation poses greater challenges to the secure operation of power grids. This paper proposes WindPower-SAFusion, an improved Informer-based model for wind power forecasting. The proposed model optimizes long-sequence modeling from three aspects. First, ProbSparse self-attention is adopted to reduce the computational complexity from O(L2) to O(LlogL). Second, a convolutional distillation encoder is introduced to compress the input sequence and highlight key temporal features. Third, a multivariate fusion and recursive multi-step forecasting framework is constructed. Using historical power and wind speed information, experiments are conducted on measured data from the Daliang Wind Farm in Guazhou, Gansu Province, China. The results show that the proposed model significantly outperforms several mainstream forecasting models in 1-day, 3-day, and 7-day forecasting tasks. Ablation experiments further demonstrate that each core module plays a critical role in improving forecasting accuracy and generalization performance. Therefore, the proposed method provides a technically feasible solution with promising engineering application potential for power grid dispatching and wind power management.