关键词: 风功率预测, 时空网络, 图注意力网络, 样本卷积和交互网络, 门控循环网络, 时间序列

Abstract: Accurate guidance and a foundation for decision-making in the wind power energy industry can be provided by accurate wind power prediction. However, the traditional focus on transforming the wind power prediction problem into a time series prediction problem, with a disregard for the spatial information among turbines, was predominantly maintained by conventional modeling methods. Therefore, a deep spatio-temporal network model for multi-time step wind power prediction was introduced. An encoder-decoder architecture was employed by the model. Firstly, a map was constructed based on historical power data by the encoder, and unit features that integrated spatial information from the wind farm were extracted using the Graph Attention Network (GAT). Secondly, characteristics from the input data were extracted by the Gated Recurrent Unit (GRU), capturing the temporal aspects of the units to acquire the temporal characteristics of wind energy. Finally, predictions for future wind power over multiple time steps were generated by the Sample Convolution and Interaction Network (SCINet), which integrated spatio-temporal features at different time scales. Experimental results indicate that the Mean Absolute Error (MAE) index of the model's predictions is as low as 42.38, showcasing a 4.25% improvement compared to the Bi-GRU (Bidirectional Gated Recurrent Unit). Moreover, the Root Mean Square Error (RMSE) index is as low as 42.71, outperforming the Autoformer by 8.70%. The advantages are validated by experiments, presenting an innovative approach for precise wind power prediction in future scenarios.

Key words: wind power prediction, spatio-temporal network, Graph Attention Network(GAT), Sample Convolution and Interaction Network (SCINet), Gated Recurrent Unit(GRU), time series