Abstract: 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 between the predictor and certain elements, such as wind and specific humidity, is not clear. The numerical prediction products of operational meso-scale model (MM5V3) in Beijing Meteorological Service and observations 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 MM5V3 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 results 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℃ as accurate for 2 m temperature, the forecast accuracy of SVM-release, MOS-release, MM5V3 model and T213 model are 66.5%, 62.2%, 58.8% and 2.5%, respectively. Forecast accuracy of 10 m wind u, v components are defined as the percentage of forecast with absolute deviations within 1 m/s, thus the forecast accuracy of SVM-release are 77.6% and 76.7%, forecast accuracy of MOS-release are 75.8% and 73.7%, forecast accuracy of MM5V3 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 MM5V3 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.