Short-term wind power forecasting in complex terrain based on spatiotemporal enhanced deep correction network

Accurate short-term wind power forecasting plays a critical role in maintaining the stability and economic dispatch of power systems. However, accurate short-term wind power forecasting in complex terrain remains a significant challenge due to strong spatial heterogeneity and temporal variability. To address these issues, we propose a Spatio-Temporal Enhanced Deep Correction Network (ST-EDCNet) that integrates a hierarchical spatiotemporal collaborative correction (HSCC) module, a spatial–semantic graph module, and a hybrid recurrent architecture. The HSCC module effectively corrects missing values, while the spatial–semantic graph module captures spatial and semantic relationships among turbines and corrects the numerical weather prediction (NWP) wind speed data. Additionally, a hybrid model combining bidirectional gated recurrent units (BiGRU) and long short-term memory networks (LSTM) with a cross-attention mechanism is used for ensemble forecasting. Experiments conducted on the Wind Farm in Sichuan Province with complex terrain demonstrate that ST-EDCNet reduces mean absolute error (MAE), root mean square error (RMSE) by 37.2% and 27.1%, respectively, and improves coefficient of determination (R2) by 10.4% compared to baseline models. The results indicate the robustness and generalization potential of our approach in complex wind farm scenarios.