Multisite Short-term Photovoltaic Prediction Based on Spatiotemporal Convolutional Neural Network

Accurate prediction of photovoltaic (PV) power generation is crucial for the smart grid and renewable energy market. In order to improve the prediction accuracy of PV power generation, a short-term photovoltaic prediction model based on spatiotemporal convolutional neural network (STCNN) is proposed. In order to capture the spatial and temporal characteristics of multi site photovoltaic power generation in one-dimensional space, a greedy adjacent algorithm (GAA) sorting method is introduced to serialize the information of two-dimensional photovoltaic stations, while keeping local photovoltaic stations adjacent to each other. The PV data is preprocessed into a spatiotemporal matrix that captures spatiotemporal correlation, facilitating the learning of convolutional neural networks. Extensive experiments are conducted on multi site photovoltaic power generation in three typical cities (Xiamen, Quanzhou, and Fuzhou). The results showed that the proposed STCNN model had an average absolute percentage error of 4.6%, 4.8%, and 5.3% for Xiamen, Quanzhou, and Fuzhou within the 6-hour prediction range of a single photovoltaic site, respectively. Compared with traditional methods, the prediction accuracy improved by 33%. When multiple photovoltaic sites are aggregated, the STCNN model proposed can reduce the error by 40% compared to existing methods.