Deterministic and probabilistic forecasting of photovoltaic power based on deep convolutional neural network

• Published in: Energy conversion and management Vol. 153; pp. 409 - 422
• Main Authors: Wang, Huaizhi, Yi, Haiyan, Peng, Jianchun, Wang, Guibin, Liu, Yitao, Jiang, Hui, Liu, Wenxin
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 01-12-2017; Elsevier Science Ltd
• Subjects: Artificial neural networks; Case studies; Clean energy; Deep convolutional neural network (DCNN); Deep learning; Electric power; Energy; Farms; Feature extraction; Forecasting; Mathematical models; Neural networks; Photovoltaic cells; Photovoltaic power forecasting; Photovoltaics; Probabilistic methods; Quantile regression; Solar cells; Spine; Statistical analysis; Wavelet decomposition; Wavelet transforms; Deep learning; Photovoltaic power forecasting; Quantile regression; Deep convolutional neural network (DCNN); Wavelet decomposition
• ISSN: 0196-8904; 1879-2227
• DOI: 10.1016/j.enconman.2017.10.008

•Deep convolutional neural network is used to extract the features in PV power data.•A deterministic model for PV power forecasting based on WT and DCNN is proposed.•Quantile regression based probabilistic PV power forecasting model is formulated.•The proposed forecasting models have been verified on PV power dataset from Belgium.
The penetration of photovoltaic (PV) energy into modern electric power and energy systems has been gradually increased in recent years due to its benefits of being abundant, inexhaustible and clean. In order to reduce the negative impacts of PV energy on electric power and energy systems, advanced forecasting approach with high-accuracy is a pressing need. Aimed at this, a novel hybrid method for deterministic PV power forecasting based on wavelet transform (WT) and deep convolutional neural network (DCNN) is firstly proposed in this paper. WT is used to decompose the original signal into several frequency series. Each frequency has better outlines and behaviors. DCNN is employed to extract the nonlinear features and invariant structures exhibited in each frequency. Then, a probabilistic PV power forecasting model that combines the proposed deterministic method and spine quantile regression (QR) is originally developed to statistically evaluate the probabilistic information in PV power data. The proposed deterministic and probabilistic forecasting methods are applied to real PV data series collected from PV farms in Belgium. Numerical results presented in the case studies demonstrate that the proposed methods exhibit the ability of improving forecasting accuracies in terms of seasons and various prediction horizons, when compared to conventional forecasting models.