A wind farm cluster power prediction method based on globally aware dynamic spatiotemporal graph continual learning framework

With the large-scale growth of wind power installed capacity, improving the power prediction accuracy of wind farm cluster (WFC) is essential for stable grid operation and efficient renewable energy use. In practical applications, traditional methods rely on static graphs with fixed turbine nodes and edges, failing to capture temporal changes due to weather and other factors, which leads to low prediction accuracy. To address this issue, we propose a Globally Aware Dynamic Spatiotemporal Graph Continual Learning (GADSG-CL) framework for WFC power prediction. Firstly, spatiotemporal correlation coefficients among wind turbines are computed using wind speed data. Subsequently, graph embedding and clustering algorithms are employed to derive multiple global information graphs that encapsulate the key features of WFC. Then, an Adaptive Temporal Label Smoothing (ATLS) method is used to suppress the clustering label noise by combining the sliding window statistical distribution with the majority voting mechanism. Finally, a spatiotemporal graph neural network incorporating a continual learning strategy is constructed to effectively adapt to dynamic graph inputs and prevent the model from forgetting historical dynamic features. Experiments on three benchmark datasets show that GADSG-CL reduces RMSE by up to 17.8% and MAE by 24.0% in mid-term forecasting, and increases R2 by about 30%. On short-term tasks, it maintains R2 values above 0.92, indicating improved accuracy and robustness.