A novel combined multi-task learning and Gaussian process regression model for the prediction of multi-timescale and multi-component of solar radiation

• Published in: Journal of cleaner production Vol. 284; p. 124710
• Main Authors: Zhou, Yong, Liu, Yanfeng, Wang, Dengjia, De, Gejirifu, Li, Yong, Liu, Xiaojun, Wang, Yingying
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.02.2021
• Subjects: accuracy; bias; China; diffuse solar radiation; Gaussian process regression; Global and diffuse solar radiation; information; learning; meteorological data; Multi-component; Multi-task learning; Multi-time scale; normal distribution; prediction; regression analysis; China; Multi-time scale; Gaussian process regression; Multi-task learning; Multi-component; Global and diffuse solar radiation
• ISSN: 0959-6526; 1879-1786
• DOI: 10.1016/j.jclepro.2020.124710

A novel combined multi-task learning and Gaussian process regression (MTGPR) model is proposed to predict the multi-time scale (daily and monthly mean daily) and multi-component (global and diffuse) solar radiation simultaneously. Compared to conventional Gaussian process regression (GPR) which can only be used for specific solar radiation component prediction on a specific timescale, the MTGPR can utilize the correlated information between different tasks to improve the model generalization and accuracy. Meteorological data from ten stations in China were used to train and validate the GPR and MTGPR models for daily global, monthly mean daily global, daily diffuse and monthly mean daily diffuse solar radiation prediction. The results showed that the GPR and MTGPR models are highly accurate in estimating daily and monthly mean daily solar radiation with coefficient of determination (R2), root mean square error (RMSE), relative root mean square error (rRMSE) and mean bias error (MBE) of GPR in ranges of 0.4623–0.9892, 0.5542–4.1591 MJm−2d−1, 4.70–39.75% and −1.1750–1.5347 MJm−2d−1, respectively. Because the MTGPR learned the intercorrelated information between different tasks, compared to GPR models, the MTGPR models performed better. For daily prediction, the average R2, RMSE and rRMSE of the MTGPR improved by 0.19–0.48%, 0.57–0.65% and 0.51–0.52%, respectively. In terms of monthly mean daily prediction, the corresponding values of MTGPR improved by 2.62–2.65%, 5.50–12.07% and 5.21–12.08%, respectively. This paper provides a compact guide for the simultaneous prediction of combined parameters. •A novel combined multi-task learning and gaussian process regression models were developed.•MTGPR and GPR performed better in predicting GSR than in predicting DSR.•MTGPR and GPR yielded higher accuracy for daily basis than for monthly basis.•MTGPR performed better than GPR by learning correlated information between tasks.