부산대학교 도서관

One major challenge in the development of solar photovoltaic (PV) systems is their inherent intermittency due to high dependability on meteorological and weather conditions. To address this challenge, providing full statistics information for the look-ahead times can be a potential solution. This paper establishes a two-stage probabilistic framework to predict full statistics information for PV power. In the first stage, a combined deep neural network is designed to capture full spatial features by a convolutional neural network (CNN), while temporal features are realized using a gated recurrent unit (GRU). In the second stage, a probability density estimation (PDF) is developed using nonparametric smooth band limit maximum likelihood (NSBML) PDF estimator to extract full information about the predicted PV power generation. The numerical results on a PV power generation dataset with a 5-min time resolution demonstrate the effectiveness and superiority of the proposed framework by comparing with several state-of-the-art models and one PDF estimator (kernel density estimator (KDE)).