Learning dynamic inter-farm dependencies for wind power forecasting via adaptive sparse graph attention network

Accurate wind power forecasting (WPF) is crucial for the stable operation and economic dispatch of power systems. Forecasting for multiple, geographically distributed wind farms is challenging due to complex and dynamic inter-farm dependencies. Existing models often rely on static relationships or struggle to adaptively integrate diverse information sources. This paper proposes an inter-farm adaptive sparse graph attention network (IF-ASGAT) to address these limitations. IF-ASGAT dynamically learns inter-farm relationships by constructing a sparse, time-varying adjacency matrix based on multivariate time series. It then employs a sparse graph attention mechanism to selectively aggregate information from the most relevant neighboring farms. The model further integrates processed spatiotemporal features with future numerical weather prediction (NWP) data for the target wind farm through a feature fusion module. Rigorous experiments on a real-world dataset of 18 wind farms show that IF-ASGAT achieves statistically significant outperformance against a wide range of baselines, including recent GNN and Transformer-based models. A comprehensive ablation study validates the indispensable roles of each modules. Furthermore, interpretability analyses highlight the model’s capability to capture physically meaningful dependencies, demonstrating that its adaptive sparsity mechanism enhances both predictive performance and computational efficiency.