What question did this study set out to answer?

This study aims to enhance the accuracy of monthly GHI predictions for solar energy optimization in China.

February 8, 2026

An enhanced monthly GHI prediction framework for China using a hybrid dynamical—machine learning approach

Key Points

This study aims to enhance the accuracy of monthly GHI predictions for solar energy optimization in China.
Developed a hybrid framework combining numerical climate models and machine learning techniques.
Utilized models like Decision Tree, Random Forest, and Naive Bayes for predictions.
Employed ERA5 reanalysis and JMA/MRI-CPS3 datasets from 1991 to 2020 for data analysis.
Applied Empirical Orthogonal Function decomposition and correlation analysis for identifying predictors.
Decision Tree model achieved the best performance with RMSE of 49.78 MJ/m² and MAE of 38.67 MJ/m².
Ensemble Mean slightly underperformed but remained more robust than other models.
Predictions exhibited relative errors within 10% compared to independent observations.
Higher accuracy found in northern stations (e.g., Ruoqiang, Harbin) versus southern ones (e.g., Fuzhou, Kunming).
Winter predictions were more accurate than summer predictions, showcasing seasonal variations.

Abstract

Accurate monthly predictions of global horizontal irradiance (GHI) are essential for optimizing solar energy integration and ensuring stable power generation. This study proposes a novel hybrid framework for monthly GHI prediction over China that combines a numerical climate prediction model (JMA/MRI-CPS3) with Linear Regression and four machine learning techniques, including Decision Tree, Random Forest, Gradient Boosting Regression Trees, and Naive Bayes. The framework extracts statistical relationships between atmospheric circulation, external forcing fields, and GHI variability using ERA5 reanalysis and JMA/MRI-CPS3 hindcast datasets from February 1991 to January 2020. Empirical Orthogonal Function decomposition and correlation analysis are applied to identify the dominant spatial–temporal modes of GHI variability and their associated predictors. The predictive performance of each individual model, as well as their Ensemble Mean, is systematically evaluated. Among the individual models, the Decision Tree exhibits the best overall performance, with RMSE = 49.78 MJ/m², MAE = 38.67 MJ/m², and SCC = 0.86. The Ensemble Mean yields slightly lower skill (RMSE = 50.45 MJ/m², MAE = 39.47 MJ/m², SCC = 0.85) but outperforms the remaining individual models and demonstrates greater robustness, producing predictions highly consistent with those of the Decision Tree. Model validation against independent observations from February 2022 to January 2024 shows good overall agreement across most regions of China, with relative errors generally within 10%. Further station-based analysis reveals clear regional and seasonal contrasts in prediction skills. Predictions at the two northern stations (Ruoqiang and Harbin) show higher accuracy, with correlation coefficients of 0.97 and 0.95, compared to the two southern stations (Fuzhou and Kunming), where correlations decrease to 0.75 and 0.70. In addition, winter predictions exhibit the highest accuracy, while summer predictions show much larger errors. Overall, this hybrid framework appears to effectively capture the nonlinear interactions among key meteorological drivers of GHI, showing promising potential for reliable monthly solar radiation prediction.

Mark Helpful

Bookmark

Relay