NWP Feature Selection and GCN-based Ultra-short-term Wind Farm Cluster Power Forecasting Method

Abstract

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 have high spatial correlation, from whose power and numerical weather prediction (NWP) data the graph convolutional neural network (GCN) can well extract spatio-temporal features. However, existing GCN-based methods for wind power forecasting have not considered the redundant information and noisy data contained in NWPs which may also be extracted by GCN, thus leading to many problems such as high model complexity and computational cost, suboptimal model training results and decrease in prediction accuracy. Focusing on this problem, this paper selects the optimal feature subsets from the available NWPs of wind farm cluster using maximum relevance minimum redundancy (MRMR) algorithm based on mutual information (MI) theory. Cross-validation is applied to ensure that the selected features maximize the valuable information of the NWPs while minimizing the redundant information and noisy data contained. The simulation results show that selecting fewer features can make errors smaller than the state-of-the-art deep learning models and reduce the computational cost under the premise of ensuring the prediction accuracy.