Kernel method is an effective method to solve the non-linear model analysis and also a research focus in the current pattern recognition community. The selection of kernel functions plays an important role in the performance of kernel methods. The Support Vector Regression (SVR) had provided higher performance than traditional learning machines and had been widely applied in real-world regression problems and nonlinear function estimation problems. In view of the regression performance, SVR is affected largely by the selected kernel function. A model of SVR based on the Bessel kernel function of the first kind were put forward and given the implementations with Rand LibSVM. Eight data sets in the database of UCI and 4 common kernel functions were selected for the experiment, the Mean Square Error and determination coefficient R2 were used as the performance evaluation index. The experimental results show that Bessel kernel function of first kind has higher prediction accuracy and stronger generalization ability in SVR, which provides references for the kernel functions selection of SVR.