An Integrated AMPSO-CLSTM Model for Photovoltaic Power Generation Prediction

Abstract

To address the challenges of volatile and intermittent nature in photovoltaic power (PV) generation forecasting, a new convolutional long short-term memory network (CLSTM) prediction model optimized by adaptive mutation particle swarm optimization (AMPSO) is proposed. In this model, the local sensing ability of the convolutional kernels in the CNN is used to extract high-dimensional features from the variable influential factors of PV power generation, and a mapping between time series data and PV is established by the memory ability of the gate control unit in LSTM. The AMPSO algorithm is introduced to optimize the network structure and weights of CLSTM simultaneously. The performance of the model is verified by two different two data sets. The results show that compared with that of the CLSTM, Auto-LSTM, LSTM and recurrent neural network models, the root mean square error (RMSE) of the AMPSO-CLSTM model decreases by 1.92–6.53% and 6.23–31.10%, the mean absolute error (MAE) decreases by 6.92–16.87% and 11.71–48.84%, and the mean absolute percentage error (MAPE) decreases by 13.24–31.75% and 12.22–49.00%, respectively. Compared with those of the CLSTM model, the number of channels in the convolutional layer of the AMPSO-CLSTM is reduced by 51.76–71.09% and 61.72–86.72%, respectively, and the number of hidden neurons in LSTM is reduced by 32–60% and 53–84%, respectively.