Due to the complexity and variability of atmospheric and vacuum distillation process, the coupling between process variables is serious, and the direct modeling will increase the difficulty of problem analysis. In order to improve the performance of the model, KPCA algorithm was used to select the variables of the model, and then the processed data were used as the input of the Gaussian process regression (GPR) model, and KPCA?GPR was used to establish the estimation model of the gasoline dry point on the atmospheric tower roof. The method solves the strong nonlinear correlation between different variables, and has the advantages of flexible nonparametric generalization and super?parameter adaptive adjustment. By calculating the empirical confidence interval, not only can the dry point of gasoline be predicted and estimated, but also can do the probability interpretation. The simulation results show that the KPCA?GPR method achieves better estimation results.

The dry point of gasoline on the top of atmospheric tower is closely related to product quality, but it is difficult to measure the gasoline dry point online, and the soft sensor is a technical way to solve the estimation and control prediction of such variables. Due to the complexity of atmospheric and vacuum distillation process, the correlation between the variables increases. In this paper, sparse principal component analysis (SPCA) was introduced into kernel principal component analysis（KPCA） algorithm, and the input variables of the model were selected by sparse kernel principal component analysis（SKPCA） algorithm. The nonlinear dimensionality reduction between data was realized, the principal component structure was simplified, and the sparsity of principal component variables was increased. The selected sparse principal components were used as the input of the least squares support vector machine (LSSVM), and the soft sensor prediction model for the top dry point of atmospheric tower was established. The simulation results show that the SKPCA?LSSVM model has higher prediction accuracy and superior model performance compared with the traditional PCA?LSSVM and KPCA?LSSVM methods.