A study of Two-Channel CNN-BiLSTM Photovoltaic Power Prediction Based on Attention Mechanism

With the increasing proportion of photovoltaic (PV) energy in the power system, accurate PV power prediction is crucial to ensure the stable operation of the power grid, enhance the power system’s ability to consume new energy sources, and reduce the demand for spare capacity due to power fluctuations. In this paper, a two-channel convolutional neural network - bi-directional short-term and long-term memory network (CNN - BiLSTM) based on attention mechanism is proposed as a PV power prediction model. The accuracy of PV power prediction is effectively improved by processing meteorological data and historical power data separately through the dual-channel structure and weighting the important features using the attention mechanism. After validation and analysis, the model is better than the existing RNN-LSTM model in terms of prediction error (MAE) and root mean square error (RMSE), and has higher accuracy and reliability than the traditional prediction model, which can provide a more accurate decision-making basis for the power dispatch department, help the power grid to realize a more reasonable power allocation and dispatch, and enhance the economic benefits of PV power generation.