Regional probabilistic forecasting of photovoltaic generation without sufficient historical data based on improved convolutional neural network

A large number of forecasting technologies are carried out on the basis of sufficient historical data, but the newly built PV plants have not been put into operation, and there is not enough historical data to forecast the power of PV plants, the existing research cannot accurately forecast the power of photovoltaic plants containing the newly built PV plants. There are two main innovations in the paper. Firstly, in the case that deep learning cannot be carried out without historical data of new PV plants, Pearson coefficients are used to analyze the correlation among PV plants and construct historical data of new photovoltaic plants. Secondly, on this basis, an improved CNN-LSTM model is proposed, which takes the form of parallel convolution-pool layers in CNN, so that the features of each plant, all plants, and time characteristics in the region can be extracted completely. Finally, the network can be fine-tuned and corrected after obtaining the actual power data. Further, the forecasting accuracy of PV power was improved, and the training time was reduced. Based on data from five PV plants in Australia, the effectiveness of the method proposed in this paper is verified.