Capturing Spatial Influence in Wind Prediction With a Graph Convolutional Neural Network

Abstract

Nowadays, wind power is playing a significant role in power systems; it is necessary to improve the prediction accuracy, which will help make better use of wind sources. The existing neural network methods, such as recurrent neural network (RNN), have been widely used in wind prediction; however, RNN models only consider the dynamic change of temporal conditions and ignore the spatial correlation. In this work, we combine the graph convolutional neural (GCN) with the gated recurrent unit (GRU) to do prediction on simulated and real wind speed and wind power data sets. The improvements of prediction results by GCN in all wind speed experiments show its ability to capture spatial dependence and improve prediction accuracy. Although the GCN does not perform well in short-term wind power prediction as the change of wind power data is not so smooth due to the limitation of turbine operation, the results of long-term prediction still prove the performance of GCN.