基于机器学习方法的瓜达尔港气温和风速预报产品订正

Temperature and wind speed prediction correction for Gwadar Port based on machine learning method

摘要:

本文分别使用支持向量机（SVM）、梯度提升回归（GBR）和随机森林（RDF）三种机器学习订正算法，对中亚区域快速更新多尺度分析和预报系统(RMAPS-CA)在瓜达尔港的气温与风速预报（2020年6月1日至2021年9月30日）进行模型训练及订正。结果表明：对于00:00UTC和12:00UTC起报的气温，SVM、GBR和RDF将预报结果的均方根误差减少了54%以上，相关系数提高到0.85以上；对于00:00UTC和12:00UTC起报的10 m风速，SVM、GBR和RDF将预报结果的均方根误差减少了63%以上，相关系数提高到0.68以上。三种机器学习订正算法中，RDF算法对近地面气温预报的订正效果最好，GBR算法则对10 m风速预报的订正效果最好。

Abstract:

In this study, we utilized the temperature and wind speed forecast products from the Rapid-update Multi-scale Analysis and Forecasting System in Central Asia (RMAPS-CA), along with weather station data for Gwadar Port, spanning from June 1, 2020, to June 30, 2021, to form our training dataset. Three distinct machine learning models—Support Vector Machine (SVM), Gradient Boosting Regression (GBR), and Random Forest (RDF) were applied to correct the forecast products of RMAPS-CA for Gwadar Port's near-surface temperature and 10 m wind speed from July 1 to September 30, 2021. The results indicate that the SVM, GBR, and RDF model has led to a notable reduction in the root mean square error (RMSE) for RMAPS-CA forecasted T-2m initialized at 00:00 UTC and 12:00 UTC, with a decrease more than 54%. Concurrently, there was a substantial increase in the correlation coefficient, which exceeded 0.85. In terms of the 10 m wind speed forecast initialized at 00:00 UTC and 12:00 UTC, the RMSE was reduced by more than 63%, and the correlation coefficient increased to above 0.68. The three machine learning models exhibited significant improvements in the accuracy of T-2m and wind speed forecasts for the Gwadar Port station. Notably, the RDF model demonstrated the most effective corrections for T-2m forecasts, while the GBR model outperformed in the correction of 10 m wind speed forecasts. These machine learning models hold considerable potential for the refinement of numerical weather prediction products in the Central Asian region.