Enhancing Short-Term Photovoltaic Cluster Power Forecasting via Spatiotemporal Generalized Weather Patterns Recognition and Dynamic Graph Convolutional Neural Networks

Due to the wide geographical coverage of the PV cluster, the power stations in it are easy to be in different weather types, resulting in varying output characteristics. However, because the cluster-scale weather patterns do not exist physically, it is difficult to comprehensively consider the different weather types of the power stations in the cluster and classify the patterns to improve the forecasting ability of the model. There is still a blank in this field at present. To fill this gap, this paper proposes a short-term PV clusters power forecasting method based on spatiotemporal generalized weather patterns recognition and dynamic graph convolutional neural networks. Firstly, the irradiance at the cluster scale is defined, and the similarity between forecasted and historical measured irradiance is calculated to identify similar daily power. Secondly, the Gaussian Mixture Model clustering algorithm is applied to classify the historical power data of the PV cluster, and the future weather patterns of the cluster are forecasted based on the similar daily power and the classification model. Finally, the short-term power forecasting model of dynamic graph convolutional neural network is constructed to capture the dynamic spatio-temporal correlations in the cluster. The validity of the proposed method is verified by the data of a PV cluster in northern China.