Monthly-Scale Global Horizontal Irradiance Forecasting Combined with Dynamics and Machine Learning
Liu Wenjing
Wang Chuanhui
Zhong Yiming
Shi Yongle
Yan Xiaojing
Abstract：In the quest to establish a monthly-scale forecasting methodology for solar radiation in China,this paper analyzes monthly Global Horizontal Irradiance(GHI)data from 53 national stations spanning 1979 to 2020,ERA5 reanalysis data,and historical simulation data from the MRI-CGCM model.An EOF decomposition is conducted on the mean GHI from 1979 to 2017 to identify principal spatial modes and corresponding temporal coefficients.These time coefficients and correlation analysis are used to comprehensively examine the atmospheric circulation within ERA5 reanalysis and MRI-CGCM simula-tion data.Significant regions,as determined by monthly segmentation and statistical testing,are selected as predictors.Besides,machine learning techniques are employed to develop a forecast model for the time coefficients.The time coefficients corresponding to the first three modes of GHI in each month of China are predicted,thereby enabling monthly-scale forecasting of GHI across China.The modeling phase involves a comparison between various machine learning algorithms and traditional stepwise regres-sion(SWR)methods.Evaluation results of mean bias and mean absolute deviation indicate that,in this study,the machine learning methods,specifically the Random Forest(RF),Gradient Boosting Decision Tree(GBDT),Decision Tree Regression(DTR)and K-Nearest Neighbors(KNN),outperform the con-ventional stepwise regression(SWR).RF exhibits superior performance in both mean bias and mean ab-solute deviation assessments.Further validation by the RF model for the prediction results in 2017 and 2018,including a comparison of forecast results and ACC scoring,demonstrates that the RF model has a certain forecasting capability in all months.Forecasts for the winter half-year surpass those for the summer half-year,with the forecasts in January,November,and December showing the best performance in the two-year forecasts.Regionally,forecasts for southern and eastern China prove to be more accurate than those for the northern and western regions of China,with the smallest deviations primarily located in East China and South China.
Keywords：MRI-CGCM modelmachine learningmonthly-scale forecastglobal horizontal irradi-ance(GHI)
Publication Date：2025-11-30
Online Publishing Date：2026-09-12(First online date of this platform, not the publication date of the document)
Pages：9( 99-107 )