Short-Term Wind Power Prediction Based on Variable Selection and MI-GCN Network

In addressing the challenges of insufficient accuracy and the difficulty in deeply mining feature information among variables in short-term wind power prediction using soft measurement methods, we propose a short-term wind power prediction model based on variable selection and an MI-GCN network. Firstly, the GRA algorithm is employed to select variables, determining the nodes of the GCN network and streamlining the model structure. Subsequently, the mutual information (MI) theory is utilized to calculate the mutual information values between nodes, simultaneously analyzing the proximity relationships of feature variables and obtaining the feature adjacency matrix. Finally, a novel short-term wind power prediction model (GRA-MI-GCN) is constructed using graph convolutional neural network (GCN). Through simulation experiments, it is validated that, compared to LSTM networks, LSSVM, and BPNN models, the MI-GCN network reduces the root mean square error (RMSE) by 4%, 26%, and 36%, respectively. It exhibits excellent generalization capability and higher prediction accuracy, demonstrating practical engineering application value.