Abstract: For the prediction of blast furnace gas(BFG) output in steel industry,which is difficult to be effectively predicted, an on-line prediction model is proposed based on data-filtering and least squares support vector machine(LSSVM). Herein,an improved empirical mode decomposition threshold filtering(IEMDTF) method is developed to filter the training data.And an online learning algorithm and Bayesian optimization approach are designed to establish and optimize the prediction model to reduce the modeling time and improve the prediction accuracy.The prediction results using practical BFG output data show that the proposed model keeps higher prediction precision under the condition of a few and random fluctuating data,and is more suitable for real time online prediction than other methods.