Multi-site Spatiotemporal Short-term Irradiance Forecasting Method Based on Transformer

Irradiance forecasting is of great significance for the management and optimization of photovoltaic power generation. Collaborative forecasting based on multiple stations can utilize richer feature information for prediction, thus there is hope to produce more accurate forecasting results. We use a spatiotemporal forecasting method based on transformer architectures and graph neural networks (GNNs) to model the relationships between different stations and different time steps. The temporal and spatial features obtained in the model are fused by a designed gate for the Global Horizontal Irradiance(GHI) forecasting of six irradiance measurement stations in Nevada. The experimental results show that this method yields better performance than STGCN and Graph WaveNet under various metrics.