Short-Term Prediction of PV Power Based on Hybrid CNN-BiLSTM-Attention Model and VMD

Accurate and reliable energy forecasting has become a mainstream trend for solving the energy crisis. The potential of databases in energy forecasting is explored by using big data-driven methods. In this paper, we introduce the variational modal decomposition (VMD) to determine the best model for short-term PV power forecasting by comparing the accuracy between VMD and different combinations of models. Each model uses recursive multi-step prediction at the high level and one-dimensional convolutional networks, as well as long short-term memory networks (LSTM), bidirectional long short-term memory networks (Bi-LSTM), and attention mechanisms at the low level. We evaluate the performance of each model strategy by comparing their mean squared error and mean absolute error against the actual values. The final results show that the VMD decomposition method proposed in this paper has the best prediction accuracy in the combined CNN-Bi-LSTM-Attention model.