Photovoltaic Power Generation Prediction Based on Spatiotemporal Graph Neural Network

Accurate prediction of regional power generation in renewable energy systems is crucial for grid dispatch optimization. However, photovoltaic power generation is subject to the influence of weather and other factors and has great randomness and unpredictability. Considering that the light intensity that affects photovoltaic power generation has great geographical and spatial similarity, the temporal and spatial similarity characteristics between stations can be used for correlation prediction. In order to fully explore the spatial similarity of photovoltaic system sites, similarity graphs such as geographical distance, altitude, equipment parameters and meteorological conditions between stations are constructed based on the prior knowledge of photovoltaic power generation physics and light intensity location altitude. Furthermore, spatiotemporal joint modeling and cross-graph joint convolution are performed on each similarity graph, and the results of graph convolution are aggregated and superimposed to strengthen the spatiotemporal characteristics and correlation of various influencing factors of photovoltaic power generation. The spatiotemporal joint modeling introduces a contextual gating mechanism and uses the historical data of similar photovoltaic power stations as context information to learn spatiotemporal dependencies. On the public PVOD dataset, the proposed spatiotemporal joint graph convolution prediction model based on multi-parameter similarity is significantly better than prediction models such as LSTM and STCNN. Compared with STCNN, which is also a spatiotemporal graph convolutional network, the prediction error is reduced by 20%. Ablation experiments show that the introduction of spatial correlation of different parameters has a greater impact on the prediction performance, and the model prediction performance improves with the increase of superimposed correlation parameters.