Ultra-short-term prediction method of PV power output based on the CNN–LSTM hybrid learning model driven by EWT

Abstract

This study proposes a new method for ultra-short-term prediction of photovoltaic (PV) power output using a convolutional neural network (CNN) and long short-term memory (LSTM) hybrid model driven by empirical wavelet transform (EWT) to address the intermittent and stochastic nature of PV power generation. Given the differences in the spatial and temporal distribution of features between PV sample data and meteorological conditions, a hybrid learning model for multibranch feature extraction was designed. First, the frequency band of PV output data was adaptively selected using EWT and decomposed into the amplitude modulation–frequency modulation single components with frequencies ranging from low to high. Second, data reconstruction was performed on the obtained power components to exploit the extraction ability of the two-dimensional CNN model for short-term local invariance and periodic features. Third, the combined one-dimensional CNN–LSTM model was used for the sample daily meteorological conditions to extract their spatiotemporal features, and the LSTM model was used to learn the correlation between the power data features and the predicted daily weather conditions and to obtain the corresponding component prediction results. Finally, the prediction results of each component were reconstructed to achieve the ultra-short-term prediction. Using Hangzhou Dianzi University's PV microgrid system as an example, the training and testing sets were randomly selected based on different seasons and weather. The results show that this method outperforms traditional learning models in terms of overall prediction performance. The proposed method of a hybrid deep learning model will provide a novel approach for ultra-short-term prediction of PV output.