A Spatial-temporal Neural Network for Photovoltaic Power Prediction

The access of photovoltaic (PV) systems in the distribution networks has facilitated accurate PV power prediction for energy coordination and operations planning. Due to the power quality and photoelectric consumption of distribution system, the prediction task is still challenging. This paper aims to provide a spatial-temporal network to capture appropriate dependencies between temporal historical PV data and spatial effect factors for the multiple-step situational predictions. It involves a simplified version of convolution and pooling operation used to capture local features and model spatial correlations in the input data, and an attention based bidirectional long short-term memory (BiLSTM) used to enhance that association data prediction. Experimental results demonstrate that the proposed prediction network indicates enhanced accuracy across various time intervals and exhibits superior performance in prediction stability and robustness when compared to typical prediction methods based on the actual measurements from a photovoltaic micro plant. The method excels in capturing both temporal and spatial dependencies, making it a valuable tool for energy planning and operations in distribution networks.