An intelligent hybrid wavelet-adversarial deep model for accurate prediction of solar power generation. (November 2021)

• Record Type: Journal Article
• Authors: Meng, Fanbin; Zou, Qiqun; Zhang, Zhanying; Wang, Bo; Ma, Hengrui; Abdullah, Heba M.; Almalaq, Abdulaziz; Mohamed, Mohamed A.
• Is Part Of: Energy reports. Volume 7(2021)
• Page Start: 2155
• Page End: 2164
• Publication Date: 2021-11
• Subjects: Solar energy -- Hybrid prediction model -- Generative adversarial networks (GANs) -- Dragonfly algorithm (DA) -- Irradiance -- Wavelet transform; Power resources -- Periodicals; Energy industries -- Periodicals; Power resources; Periodicals; Electronic journals; 621.04205
• DOI: 10.1016/j.egyr.2021.04.019
• Languages: English
• ISSNs: 2352-4847

Abstract

Abstract: A novel intelligent hybrid model is proposed in this paper for accurate prediction and modeling of solar power plants using the wavelet transform package (WTP) and generative adversarial networks (GANs). In the proposed model, the wavelet transform is deployed for decomposing the solar energy signal into the sub-harmonics followed by the statistical feature selection analyses. The GAN model as a deep learning approach is proposed for learning each sub-frequency and predicting the future of the solar energy in the short-time window. Due to the high complexity of the solar irradiance data when training, an evolutionary algorithm based on dragonfly algorithm (DA) is suggested to train the generative and discriminator networks in the GAN. Moreover, a three-phase adaptive modification is suggested to enhance the search capabilities of the DA optimization. The efficiency and appropriate performance of the proposed model is examined and compared with the most successful models such as artificial neural networks (ANNs), support vector machine (SVM), time series, auto-regressive moving average (ARMA), and original GAN on several benchmarks for varied forecast time horizons. The simulation results on the datasets of two regions show the mean absolute percentage error (MAPE) of 0.0282 and 0.0262, when 1-pace forecast horizon, for the two regions which increase up to 0.0531 and 0.0631 for 6-pace forecast horizon. Moreover, the root mean absolute error (RMSE) is 0.0473 and 0.0479 for the two regions at 1-pace forecast horizon which increased up to 0.0895 and 0.0946 for 6-pace forecast horizon. These results show the more precise performance of the proposed forecast deep learning as well as the more optimal performance of the modified DA over the other algorithms shown in the results.