Abstract: The design of gasoline blending made a linear transformation of the set of variables comprising component octane number, component distillation process data, component proportion of blending gasoline,and additive quantity.With a production-based data LSSVM and ν-SVR and RBFNN machine learning methods the prediction model for octane number was established, and the design was computerized for directing gasoline blending in a petroleum refinery.The experimental results showed that the mean of value of absolute error was 0.227RON,and the maximun absolute error was 0.480RON.The experiments with the absolute error within 0.3RON accounted for 76.2% of the total experiments while those with absolute error within 0.4RON accounts for 90.5% of total experiments.