Membrane Contamination Prediction Based on BiLSTM and Weight Combination Strategy

Abstract:

Aiming at the membrane contamination problem that is very likely to occur when recovering proteins from gluten processing wastewater by membrane separation method, a weight combination model based on bi-directional long shortterm memory (BiLSTM) was proposed for the prediction of membrane contamination status. Taking the 14 relevant variables collected from the pilot production line of gluten processing wastewater extraction and recycling as inputs, and the changes in membrane flux as outputs, four baseline models were established: support vector machine model (SVM), back propagation neural network model (BP), random forest model (RF), generalized regression neural network (GRNN), together with one given model: BiLSTM model. The weights of the baseline model and the given model were calculated by the inverse error method to construct the weight combination prediction model. Finally, the prediction performance of the single model and the weight combination model was analyzed by using the coefficient of determination R2 and the mean square error (MSE) as the evaluation indexes. The results showed that the weight combination model was able to synthesize the advantages of the singleitem model and significantly outperformed the single-item model in terms of performance. Among them, the BP+BiLSTM+RF model had a high R2 of 0.9906 with high fitting accuracy and MSE of 1.004L2/(h2·m4), which was the lowest among all models. Compared with BP, BiLSTM and RF single-item models, the reduction was 46.05%, 67.24% and 50.81%, respectively. The developed weight combination model can be used for accurate prediction of membrane contamination during protein recovery treatment of gluten processing wastewater.