A short-term probabilistic photovoltaic power prediction method based on feature selection and improved LSTM neural network

With the increase of solar photovoltaic(PV) penetration in power system, the impact of random fluctuation of PV power on the secure operation of power grid becomes more and more serious. An efficient PV forecasting approach is proposed to accurately quantify the variability and uncertainty of the power production from PV systems. This study proposes a classification method of weather types based on cloud cover and visibility. A PV power forecasting model is proposed, based on various meteorological data including cloud cover and visibility and in order to make the model show better performance, Maximal Information Coefficient(MIC) is used to select the feature variables. Coupled Input and Forget Gate(CIFG) network is proposed to minimize structure without significantly decreasing the prediction accuracy. Furthermore, a new hybrid method combining Quantile Regression(QR) and CIFG network is proposed to predict the conditional quantile of PV output. Afterward, Kernel Density Estimation(KDE) method is used to estimate PV output probabilistic density function(PDF) according to these conditional quantiles of PV output. The effectiveness and high reliability of the proposed forecasting model are demonstrated through several other forecasting methods, and a significant improvement in PV power prediction is observed.