The dry point of gasoline is difficult to be measured in real time. A large number of data samples need to be extracted to test the quality of each section of the oil. In order to solve this problem, predictive control was carried out by establishing a soft sensor model. The least squares support vector machine model is too sensitive to outliers, which is easy to affect the prediction accuracy. By establishing the weighted least squares support vector machine model (WLSSVM), the fitting error is weighted, which weakens the influence of outliers on the model and improves the anti?interference ability of the model. The improved WLSSVM was applied to the prediction of gasoline dry point. The results show that the maximum absolute error of the improved WLSSVM is 11.65% lower than that of the least squares support vector machine model, and its prediction performance and robustness have obvious advantages.