Predicting membrane cleaning effectiveness in a full-scale water treatment plant using an artificial neural network model

Membrane fouling is the primary operational challenge of membrane technologies in full-scale treatment facilities. However, quantification of membrane fouling is challenging because membrane operation is governed by a complex combination of uncertain and non-linear process parameters. Data-driven models using machine learning (ML) can be an efficient tool to characterize the membrane fouling process since they can deal with complicated datasets that include many parameters with uncertainty and non-linearity. In the current study, an artificial neural network (ANN) was trained on an approximately one-year-long dataset collected from four membrane racks used in a drinking water treatment facility. The ANN was used to predict the membrane specific flux (permeance) and recovery in specific flux after chemically-enhanced backwash (CEB) events using eleven input variables including membrane cleaning protocols. The ANN model resulted in high descriptive accuracy, with R2 values of approximately 0.97. A feature importance analysis demonstrated that the output variables were significantly controlled by the drop in specific flux in the filtration cycle occurring before a CEB, and the specific flux measured immediately before a CEB. Our results show that this ANN model is an effective tool to inform water treatment operators about the expected specific flux and recovery that can be achieved after a CEB cleaning event.