Adaptive Feature Selection and GCN With Optimal Graph Structure-Based Ultra-Short-Term Wind Farm Cluster Power Forecasting Method

Cluster-level wind power forecasting is of great significance for the centralized integration of wind power into the grid. Studies have shown that adjacent wind farms exhibit high spatial-temporal correlation. As an extension of the convolutional neural networks (CNN), the graph convolutional neural networks (GCN) can effectively extract spatial-temporal features from the power and numerical weather prediction (NWP) data of adjacent wind farms. However, the strong correlation among NWP data from various wind farms within the same region inevitably leads to higher redundancy. Directly modeling all the wind farms of the cluster as a graph input for GCN would result in increased complexity and computational costs of the prediction model, thereby affecting the performance and accuracy of the prediction model. Therefore, it is necessary to perform feature selection on the wind farm cluster. To address the issue of manually determining the optimal number of features in the traditional maximum relevance minimum redundancy (MRMR) algorithm through cross-validation, an adaptive MRMR algorithm is proposed by introducing conditional mutual information. This algorithm automatically determines the optimal number of features in the feature subset. The optimal feature subsets obtained are used to construct an optimal graph structure as input for GCN in wind farm cluster power forecasting. Simulation results demonstrate that the proposed method has lower data and computational costs while exhibiting outstanding performance in improving power prediction accuracy.