Abstract: On the basis of particle swarm optimization（PSO）algorithm and support vector machine（SVM），the PSO algorithm with last out mechanism was applied to optimize the parameters of SVM. Then，a mutant ICPSO-LSSVM model for predicting soft sensor of the evaporation process in alumina production was constructed and compared with PSO-LSSVM and LSSVM models. Results illustrated that ICPSO-LSSVM is featured with more accuracy and better generalization ability and performance，with which samples with prediction relative error less than 5% are as high as 92.5%，the maximal relative error is only 8.1%，and the MSE is 0.05153. The implement experimental results indicated that the model with more accuracy can basically realize the real-time and online estimation for output concentration.