A hybrid model with dual channel feature processing for short-term photovoltaic power prediction

Abstract

Adverse effects of random fluctuations and intermittent characteristics of solar irradiance usually hamper the proper operation of the photovoltaic power grid. It is therefore desirable to improve the accuracy of photovoltaic (PV) power prediction. In this work, PV forecasting is realized through a Bayesian optimized model which combines the long short-term memory and radial basis function neural network (BOA-LSTM-RBF). The hybrid model presents a dual channel feature processing by extracting the historical data of PV generation via long-short-term memory network (LSTM) and extracting the forecasted weather conditions via radial basis function neural network (RBF). Then the number of hidden layer neurons and the training batch size are simultaneously optimized by & the Bayesian optimization algorithm (BOA). The testing results of three stations demonstrate that, compared with other available models, the RMSE values of BOA-LSTM-RBF model decreased by 2% ∼ 17%, which has striking advantages in prediction precision and generalizability. More interestingly, high-precision PV power forecasting can be achieved even under dramatic weather changes.