Two methods were proposed to optimize the training set of chemical engineering modeling based on particle swarm optimization (PSO) with random sampling.The forecasting model was formed by integrating PSO and other machine learning arithmetic which was used to model.In method 1 the optimized training set was acquired firstly by optimizing the training set which set the goal of minimizing the mean of squared errors (mse) in cross validation

secondly by making the prediction to the validation set

and lastly by choosing the one that gave the best prediction results among the optimized training sets.In method 2 the thought of early stop in BP neural network was adopted.The optimized training set was acquired firstly by optimizing the training set which set the goal of minimizing the mse to the validation set

secondly by making the prediction to the testing set

and lastly by choosing the one that gave the best prediction results among the optimized training sets.These two methods were used to deal with the simulation data and the gasoline blending data collected from a refinery.The results showed that the methods could improve the accuracy of prediction greatly.These two methods are worth generalizing in chemical engineering modeling in the future.