Because static soft sensor modeling can not reflect the dynamic information of industrial processes，which lead to worse estimation precision and robustness.A dynamic soft sensor modeling based on least square vector machine（LS-SVM）and ARMA time series prediction modeling was presented.A static soft sensor model based on LS-SVM was established firstly，and then dynamic correction in the static model was made by using the dynamic estimation of prediction error in ARMA to improve the dynamic response characteristics.Finally，the proposed LSSVM-ARMA was used to predict the concentration of ethane in ethylene distillation.Simulation indicated that this method featured good approximation and good generalization ability as compared with LSSVM，and could be used in soft sensor.