Hybrid method for short-term photovoltaic power forecasting based on deep convolutional neural network

• Published in: IET generation, transmission & distribution Vol. 12; no. 20; pp. 4557 - 4567
• Main Authors: Zang, Haixiang, Cheng, Lilin, Ding, Tao, Cheung, Kwok W, Liang, Zhi, Wei, Zhinong, Sun, Guoqiang
• Format: Journal Article
• Language: English
• Published: The Institution of Engineering and Technology 13.11.2018
• Subjects: CNN; commonly used methods; convolution kernels; deep convolutional neural network; different frequency components; grid‐connected PV systems; historical time series; hourly timescales; hybrid method; hybrid model; inexhaustible renewable energy; load forecasting; neural nets; photovoltaic electric power; photovoltaic power systems; power engineering computing; rising energy demands; short‐term photovoltaic power forecasting; short‐term PV power forecasting; solar radiation; Special Issue: New Trends in the Planning of Distribution Network with High Penetration of Renewables and Flexible Loads; time series; variational mode decomposition; VMD; historical time series; CNN; hybrid method; deep convolutional neural network; commonly used methods; short-term photovoltaic power forecasting; short-term PV power forecasting; photovoltaic power systems; hourly timescales; time series; solar radiation; different frequency components; VMD; power engineering computing; hybrid model; inexhaustible renewable energy; variational mode decomposition; photovoltaic electric power; load forecasting; convolution kernels; grid-connected PV systems; neural nets; rising energy demands
• ISSN: 1751-8687; 1751-8695
• DOI: 10.1049/iet-gtd.2018.5847

Photovoltaic (PV) electric power has been widely employed to satisfy rising energy demands because inexhaustible renewable energy is environmentally friendly. In order to mitigate the impact caused by the uncertainty of solar radiation in grid-connected PV systems, a hybrid method based on a deep convolutional neural network (CNN) is introduced for short-term PV power forecasting. In the proposed method, different frequency components are first decomposed from the historical time series of PV power through variational mode decomposition (VMD). Then, they are constructed into a two-dimensional data form with correlations in both daily and hourly timescales that can be extracted by convolution kernels. Moreover, the time series of residue from VMD is refined into advanced features by a CNN, which could reduce the data size and be easier for further model training along with meteorological elements. The hybrid model has been verified by forecasting the output power of PV arrays with diverse capacities in various hourly timescales, which demonstrates its superiority over commonly used methods.