Adaptive Encoder-Decoder Model Considering Spatio-Temporal Features for Short-Term Power Prediction of Distributed Photovoltaic Station

Considering the impact of operation and maintenance costs and technology, there is generally a lack of sufficient meteorological observation devices within the distributed photovoltaic (PV) station group. The deviation of the collected meteorological data and the PV power data error caused by software and hardware limitations will directly lead to the reduction of model prediction accuracy. To tackle this problem, this paper proposes a short-term prediction method with adaptive spatio-temporal codec structure for distributed PV power prediction, which adapts to the prediction requirements of different data input and different weather conditions and improves the prediction accuracy. First, the Random Forest algorithm (RF) and Pearson Correlation Coefficient (PCC) are used to sort the feature importance and select the key input data. Second, a spatio-temporal feature encoder-decoder model based on Long Short Term Memory Network (LSTM) and Spatio-Temporal Attention mechanism (STA) is proposed to adapt to spatio-temporal feature mining under different weather conditions. Third, an adaptive prediction framework based on pre-fusion and post-fusion is designed to meet the needs of comprehensive feature learning under the input of different amounts of data and further improve the prediction accuracy. Comprehensive experiments have been conducted on real data from different stations in Jiangsu, China, to confirm the superior performance of the proposed model.