Spatio-Temporal Graph Neural Network and Pattern Prediction Based Ultra-Short-Term Power Forecasting of Wind Farm Cluster

Accurate and timely ultra-short-term wind farm cluster power forecasting is significant for real-time dispatch and frequency regulation of power grids. Distinguishing different types of power fluctuation patterns based on fluctuation process analysis and training prediction models separately based on pattern partitioning results, is beneficial for improving the prediction accuracy of wind farm cluster power. However, existing pattern partitioning methods have a single perspective and have not yet formed a multi-dimensional evaluation routine to quantify the fluctuation characteristics of different patterns. Furthermore, for wind farm clusters, there is a lack of consideration of the dynamic spatio-temporal relationship between adjacent wind farm stations under different power fluctuation patterns. To make up for these deficiencies, this article proposes an ultra-short-term wind farm cluster power forecasting model based on power fluctuation pattern recognition and spatio-temporal graph neural network pattern prediction. First, the extreme points are statistically analyzed, and the wind farm cluster power is divided into different fluctuation processes. Then four indicators are summarized from the time stationarity and amplitude volatility of these fluctuation processes to guide the partition of power fluctuation patterns. Finally, considering the dynamic spatio-temporal correlation between adjacent stations under various fluctuation patterns, the spatio-temporal graph neural network is exploited for model training for each fluctuation pattern. After identifying the fluctuation patterns of the wind power series in the test set, the corresponding trained model is used to obtain the final prediction results. Experiments with other benchmarks show that the proposed method is superior on real wind farm cluster power dataset.