Enhancing accuracy of membrane fouling prediction using hybrid machine learning models

Abstract

Membrane fouling significantly affects membrane performance, but cleaning and replacement schedules are often set at regular time intervals, regardless of the extent of deterioration in performance. The aim of this study is to develop an improved prediction model for membrane fouling in the seawater reverse osmosis (SWRO) process using a hybrid machine learning approach. A Kalman filter (KF) was combined with either an artificial neural network (ANN) or a support vector machine (SVM)- a family of machine learning models-in series. The performance of these integrated models was evaluated with training and testing data sets compiled from the Fujairah SWRO plant over a period of 18 months. Our findings showed that the SVM alone provided, on average, slightly better prediction of membrane resistance (an indirect indicator of membrane fouling) than a single ANN during training and testing sets. However, hybrid machine learning methods consistently outperformed any single model, with the combination of the KF and SVM exhibiting better performance than that of the KF and ANN, except for one special case in which the accuracy of a single SVM already exceeded 0.8 for both Nash-Sutcliffe model efficiency and R-2. Taken together, our results demonstrated that the hybrid machine learning approach not only enhanced the prediction ability of membrane resistance in classical fouling and machine learning models, but could also be used to adjust cleaning and replacement schedules correctly in response to progressive deterioration in membrane performance during operation.