Aiming at the problems that the power of photovoltaic power generation is greatly affected by meteorological factors, has strong volatility, and has low prediction accuracy, this paper proposes a short-term photovoltaic power generation prediction method based on the multi-scale spatiotemporal correlation of arrays. Fully considering the spatial correlation among photovoltaic arrays and the historical data characteristics at different time scales, a deep learning prediction model integrating graph convolutional networks, multi-scale convolution, and dual attention mechanisms was constructed. First, the topological associations and power propagation patterns among multiple arrays are captured through the spatial graph convolutional network. Second, a three-branch parallel convolutional structure is designed to extract multi-scale temporal features such as short-term fluctuations, medium-term trends, and long-term cycles. Third, a dual attention mechanism is introduced to achieve adaptive weighting of feature dimensions and time steps.; Finally, a multi-layer long short-term memory (LSTM) network is adopted for sequence modeling and power prediction. Verification based on the actual operation data of a large-scale photovoltaic power station for 12 consecutive months shows that, compared with hybrid deep learning baseline methods, the average absolute percentage error of the proposed method is reduced by ∼18.7% compared with the best baseline graph convolutional network (GCN)-LSTM, and the coefficient of determination R2 reaches 0.934, effectively improving the short-term photovoltaic power prediction accuracy and providing reliable technical support for power grid dispatching, energy storage configuration, and energy management.
Ma et al. (Fri,) studied this question.