DBANN: Dual-Branch Attention Neural Networks with hierarchical spatiotemporal-perception for multi-node offshore wind power forecasting

Precise forecasting of offshore wind power is the basic guarantee for ensuring power grid operation stability and optimizing energy management. Nevertheless, existing approaches struggle to balance multi-scale temporal characteristics, unstructured spatial dependencies, and intertwined spatiotemporal interactions. This study proposed the novel Dual-Branch Attention Neural Networks (DBANN) framework, including a Spatial Branch for capturing spatial correlations, a Temporal Branch for extracting hierarchical features, and a Spatiotemporal Fusion Module to integrate them. Experiments on real-world offshore wind farm data provided evidence for the proposed DBANN model's enhanced performance relative to SOTA models, achieving a 48.67 % reduction in RMSE, 40.30 % in MAE, and 52.85 % in MAPE. Ablation studies validate the contributions of multi-resolution convolutional kernels, dual-graph structures, and feature fusion mechanisms. Moreover, the parallel architecture outperforms cross-configurations, underscoring the significance of independent yet complementary extraction of the spatiotemporal feature.