Short-term photovoltaic power forecasting based on a new hybrid deep learning model incorporating transfer learning strategy

Abstract

The accurate prediction of photovoltaic (PV) power generation is an important basis for hybrid grid scheduling. With the expansion of the scale of PV power plants and the popularization of distributed PV, this study proposes a multilayer PV power generation prediction model based on transfer learning to solve the problems of the lack of data on new PV bases and the low accuracy of PV power generation prediction. The proposed model, called DRAM, concatenates a dilated convolutional neural network (DCNN) module with a bidirectional long short-term memory (BiLSTM) module, and integrates an attention mechanism. First, the processed data are input into the DCNN layer, and the dilation convolution mechanism captures the spatial features of the wide sensory field of the input data. Subsequently, the temporal characteristics between the features are extracted in the BiLSTM layer. Finally, an attention mechanism is used to strengthen the key features by assigning weights to efficiently construct the relationship between the features and output variables. In addition, the power prediction accuracy of the new PV sites was improved by transferring the pre-trained model parameters to the new PV site prediction model. In this study, the pre-training of models using data from different source domains and the correlations between these pre-trained models and the target domain were analyzed.