Forecasting Photovoltaic Power using a Convolutional Neural Network and the Wavelet Transformation

Abstract

In order to reduce the pressure caused by the precipitous depletion of natural resources and realize the sustainable development, renewable energy resource is playing a more and more significant role with an exponential growth speed in the smart grid, especially for solar energy. Given the stochastic characteristics of the Photovoltaic (PV) generation and to maintain the security and reliability of the power grid operation, finding a forecasting method with higher accuracy is a pressing need. Facing with this challenge, a novel prediction model that combines Convolutional Neural Network with Wavelet Decomposition (CNN+WD) is presented in this paper. WD technique is aimed to decompose the historical photovoltaic power data into several wavelets with different frequencies. CNN is used to train the decomposed wavelets separately and predict the future output PV power. Three other kinds of models are investigated to compare with the presented approach. The performance is respectively assessed for 15mins, 30mins, 45mins, 60mins, 90mins and 120mins ahead power forecasting in deterministic prediction. Numerical results presented in simulations show that the proposed model has good reliability and high accuracy in comparison with other models selected for this study.