Regional distributed photovoltaic power forecasting considering spatiotemporal correlation and meteorological coupling

Current distributed photovoltaic power forecasting methods typically use static graph models to capture the spatiotemporal characteristics among distributed photovoltaic power stations， but most of them do not account for the varying impact of meteorological factors on the power forecasting of different stations. To address this， this paper proposes a regional distributed photovoltaic power forecasting method that considers spatiotemporal correlation and meteorological coupling. First， based on an analysis of the output characteristics of distributed photovoltaic power stations， an adaptive graph convolutional neural network combined with a long short-term memory network （LSTM） is used to extract the spatiotemporal features of the photovoltaic output. Additionally， a neural network layer with non-shared parameters is employed to capture the coupling relationship between different photovoltaic stations and meteorological factors， enabling the forecasting of power generation across multiple stations. To reduce the error amplification caused by directly summing the predicted power of each station， a learnable weight layer is introduced into the model to obtain the total regional photovoltaic power. Finally， through comparative analysis with various forecasting models under multiple weather scenarios， the proposed method is validated for its accuracy and stability.