Objective To prevent and treat of ceramic membrane purification of membrane fouling process of TCM extracts; To explore new methods of forecasting membrane fouling degree.Methods BP neural network model was improved. Methods to fast determine the optimal number of neurons in the hidden layer and fast algorithm for optimizing the weight and threshold of BP neural network were studied. Data of 207 groups of TCM extracts were under network training and prediction.ResultsCompared with the models of multiple regression analysis, basic BP neural network and RBF neural network, the error of the improved BP neural network model was less than that of the BP neural network model, and the mean square error was only 0.0057. In addition, the improved BP neural network model performance was more stable. In the 20 random running experiments, the goal of the success rate achieved up to 95%.Conclusion The improved model has a good network performance, the fitting effect and prediction ability, and can forecast the fouling degree of membrane stably and accurately.