A novel spatiotemporal relation fused network for solar photovoltaic power prediction

This paper proposes a novel spatiotemporal relationship fusion network (STRFN) for short-term photovoltaic power generation prediction. STRFN integrates CNN for spatial feature extraction, LSTM for time-dependence capture, and an attention mechanism. Bayesian optimization is utilized to identify optimal hyperparameters, and advanced data preprocessing techniques improve input data quality. Applied to two typical PV systems, STRFN achieves a RMSE of 2.46 and 0.036, and a MAPE of 1.51% and 1.94%, significantly outperforming traditional deep-learning models. It also shows consistent superiority across four seasons, with NRMSE values of 0.90% and 0.61% for the two systems. Additionally, a one-stop prediction platform integrating data processing, model training, and result visualization is developed. The method and platform offer valuable insights for researchers and PV operators.