支持向量机非线性回归方法的气象要素预报

Prediction of Meteorological Elements Based on Nonlinear Support Vector Machine Regression Method

该文介绍了基于基本的支持向量机非线性回归方法,该方法具有解决非线性问题的能力,在数值预报解释应用技术中,对某些预报量与预报因子之间相关性不显著的要素,如风、比湿等,采用支持向量机非线性回归技术较多元回归的MOS方法更具优势;利用北京市气象局中尺度业务模式(MM5V3)的12:00(世界时)起始数值预报产品和观测资料,制作北京15个奥运场馆站点6～48 h逐3 h的气象要素释用产品。对比MM5V3模式,从均方根误差的平均减小率来看,2 m温度减小12.1%,10 m风u分量减小43.3%,10 m风u分量减小53.4%,2 m比湿减小38.2%。与同期的MOS方法预报结果相比,整体预报效果SVM略优于MOS。由此可见,支持向量机非线性回归方法解决与预报因子之间非线性相关的气象要素较好,具有较高的预报优势。

Nonlinear regression method based on basic support vector machine is introduced, which is able to solve nonlinear problems. The cross validation in this method makes it able to optimize the kernel function parameters. Therefore, in numerical weather prediction interpretation, nonlinear support vector machine regression technique is better than multi-variant MOS regression method when the linear relationship be- tween the predictor and certain elements, such as wind and specific humidity, is not clear. The numerical prediction products of operational meso-scale model （MMSV3） in Beijing Meteorological Service and obser- vations are used to make 6--48 h interpretation products with 3-hour interval of meteorological elements of 15 venues stations in Beijing. The comparison of interpretation products and MMSV3 forecast indicates that the root mean square error for 2 m temperature, 10 m wind u component, 10 m wind v component and 2 m specific humidity reduces by 12.1%, 43.3%, 53.4% and 38.2%. Compared with the prediction re- sults of MOS, 2 m temperature, 10 m wind u compoent, v component prediction results of SVM are slightly better than those of MOS, and 2 m specific humidity prediction result of SVM is better than that of MOS. Defining the forecast with deviations no more than 2~C as accurate for 2 m temperature, the forecast accuracy of SVM-release, MOS-release, MM5V3 model and T213 model are 66.5%//00, 62.2%, 58.8% and 2.5 %, respectively. Forecast accuracy of 10 m wind u, v components are defined as the percentage of fore- cast with absolute deviations within 1 m/s, thus the forecast accuracy of SVM-release are 77. 6~/oo and 76.7%, forecast accuracy of MOS-release are 75.8% and 73.7%, forecast accuracy of MMSV3 model are 54.5% and 41.1%, and forecast accuracy of T213 model are 46.9% and 34.9%. Forecast accuracy of 2 m specific humidity is defined as the percentage of forecast with absolute deviations within 2 g kg-1 , thus the forecast accuracy of SVM-release, MOS-release and MMSV3 model are 84.9%, 67.8% and 61.7%, respectively. It shows that nonlinear support vector machine regression method is good at solving nonlinear dependence between meteorological elements and predictors, and performs better than MOS.