Ultra-short-term Photovoltaic Power Prediction Based on Multi-head ProbSparse Self-attention and Long Short-term Memory

Abstract

Abstract To provide accurate predictions of photovoltaic (PV) power generation, an MHPSA-LSTM ultra-short-term multipoint PV power prediction model combining Multi-head ProbSparse self-attention (MHPSA) and long short-term memory (LSTM) network is posited. The MHPSA is first used to capture information dependencies at a distance. Secondly, the LSTM is used to enhance the local correlation. At last, a pooling layer is added after LSTM to reduce the parameters of the fully-connected layer and alleviate overfitting, thus improving the prediction accuracy. The MHPSA-LSTM model is validated on a PV plant at the Desert Knowledge Australia Solar Centre as an example, and the RMSE, MAE, and R2 of MHPSA-LSTM are 0.527, 0.264, and 0.917, respectively. MHPSA-LSTM has higher prediction accuracy compared with BP, LSTM, GRU, and CNN-LSTM.