Wavelet-based 3-phase hybrid SVR model trained with satellite-derived predictors, particle swarm optimization and maximum overlap discrete wavelet transform for solar radiation prediction

• Published in: Renewable & sustainable energy reviews Vol. 113; p. 109247
• Main Authors: Ghimire, Sujan, Deo, Ravinesh C., Raj, Nawin, Mi, Jianchun
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2019
• Subjects: Hybrid energy prediction model; Maximum overlap discrete wavelet; MODIS satellite; Remote sensing; Renewable energy exploration; Support vector regression; Support vector regression; Hybrid energy prediction model; Renewable energy exploration; MODIS satellite; Remote sensing; Maximum overlap discrete wavelet
• ISSN: 1364-0321; 1879-0690
• DOI: 10.1016/j.rser.2019.109247

The accurate prediction of global solar radiation (GSR) with remote sensing in metropolitan, regional and remote, yet solar-rich sites, is a core requisite for cleaner energy utilization, monitoring and conversion of renewable energy into usable power. Data-driven models that investigate the feasibility of solar-fueled energies, face challenges in respect to identifying their appropriate input data as such variables may not be available at all sites due to a lack of environmental monitoring system. In this paper, the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite-derived predictors are employed to train three-phase hybrid SVR model for monthly GSR prediction. Firstly, to acquire relevant model input features, MODIS variables are screened with the Particle Swarm Optimization (PSO) algorithm, and secondly, a Gaussian emulation method of sensitivity analysis is incorporated on all screened variables to ascertain their relative role in predicting GSR. To address pertinent issues of non-stationarities, PSO selected variables are decomposed with Maximum Overlap Discrete Wavelet Transformation prior to its incorporation in Support Vector Regression (SVR), constructing a three-phase PSO-W-SVR hybrid model where the hyper-parameters are acquired by evolutionary (i.e., PSO & Genetic Algorithm) and Grid Search methods. Three-phase PSO-W-SVR hybrid model is benchmarked with alternative machine learning models. Thirty-nine model scenarios are formulated: 13 without feature selection (e.g., SVR), 13 with feature selection (e.g., PSO-SVR for two-phase models) and the remainder 13 with feature selection strategy coupled with data decomposition algorithm (e.g., PSO-W-SVR leading to a three-phase model). Metrics such as skill score (RMSESS), root mean square error (RMSE), mean absolute error (MAE), Willmott’s (WI), Legates & McCabe’s (E1) and Nash–Sutcliffe coefficients (ENS) are applied to comprehensively evaluate prescribed models. Empirical results register high performance of three-phase hybrid PSO-W-SVR models, exceeding the prescribed alternative models. High predictive ability evidenced by a low RRMSE and high E1 ascertains PSO-W-SVR hybrid model as considerably favorable in its capability to be enriched by MODIS satellite-derived variables. Maximum Overlap Discrete Wavelet Transform algorithm is also seen to provide resolved patterns in satellite variables, leading to a superior performance compared to the other data-driven model. The research avers that a three-phase hybrid PSO-W-SVR model can be a viable tool to predict GSR using satellite derived data as predictors, and is particularly useful for exploration of renewable energies where satellite footprint are present but regular environmental monitoring systems may be absent. [Display omitted] •SVR (predictive model), PSO (feature selection) and maximum overlap DWT (frequency resolution wavelet model) is integrated.•Three-phase PSO-W-SVR hybrid model is trained with satellite inputs for solar energy prediction.•Hybrid PSO-W-SVR outperforms all tested data-driven models.•Hybrid PSO-W-SVR model is a potential tool for long-term solar energy exploration.