A Novel High-Resolution Prediction Model of Photovoltaic Power Generation with a Multi-Attention Mechanism

Photovoltaic power generation is influenced by weather variations, solar radiation intensity, and equipment status, which can exhibit significant fluctuations over short periods. Highresolution time series data can capture these dynamic changes more accurately, thereby improving prediction accuracy and the precision of digital twins and intelligent decision-making processes. A multi-attention mechanism is introduced to predict highresolution photovoltaic (PV) power and capture the complex nonlinear relationships within multivariate time series while considering weather factors. Consequently, the SCINet-BiLSTM-Multi-Head Attention model (Abbreviation SB-MHA) is developed. These extracted features are then fused, and the multihead attention mechanism is employed to learn the most critical factors from the data. The model is optimized using the policy gradient optimization algorithm and is compared against eight other neural network models. The results show that the root mean square error (RMSE) decreases by 16.367, 2.607, and 0.061 compared to the CNN-BiLSTM, BiLSTM-Attention, and CNN-LSTM-Attention models.