ISSN 1000-0526
CN 11-2282/P
CN 11-2282/P
Archive > Volume 51 Issue 8 > 2025,51(8):941-953. DOI:10.7519/j.issn.1000-0526.2025.010902 Prev Next
Correction Method of Shortwave Radiation Numerical Forecast in Henan Province Based on Machine Learning
- Article
- Figures
- Metrics
- Preview PDF
- Reference
- Related
- Cited by
- Materials
Abstract:
Based on the total irradiance data from 23 radiation observation stations in Henan Province in 2022 and CMA-WSP2.0 model products, in this paper, characteristic variables are selected by LASSO regression, training data sets and test data sets are established, and models are trained by machine learning methods (random forest, XGBoost and LightGBM) with the training data sets. Besides, the total irradiance forecasts by CMA-WSP2.0 model in Henan Province are corrected, and the corrected results are tested by site and region, by season and by total irradiance classification. The results are as follows. The three machine learning methods (random forest, XGBoost and LightGBM) have good correction effects. Compared with the CMA-WSP2.0 prediction results, the mean absolute error (MAE) and root mean square error (RMSE) are significantly reduced, and the 24 h accuracy rate and 24 h qualification rate are significantly improved. Among them, LightGBM has the best correction effect. The MAE decreases by 18.32-32.91 W·m-2, the reduction proportion of MAE decreases by 38%-56%, and the reduction proportion of RMSE decreases by 36%-52%. Moreover, the 24 h average accuracy and 24 h average qualification rate are raised by 7.3% and 5.7%, respectively. The results of regional statistics are consistent with those of the stations. For the five regions of Henan Province, the correction effect for western Henan is the best. The corrected deviation range of the three machine machine learning methods is narrower than that of the CMA-WSP2.0 test set, and the probability of the deviation distribution near zero is greater. Among the seasonal test results, the three machine learning methods have more significant correction effect for the winter prediction. For different total irradiance levels, the three machine learning methods can effectively improve the CMA-WSP2.0 prediction, but the correction effect tends to gradually weaken with the increase of total irradiance levels. These findings could provide a useful reference for improving the ability of total irradiance forecast in Henan Province.
Keywords:
Project Supported:
Clc Number:
Mobile website
