Predicting Membrane Fouling of Submerged Membrane Bioreactor Wastewater Treatment Plants Using Machine Learning

Abstract

Membrane fouling remains a significant challenge in the operation of membrane bioreactors (MBRs). Plant operators rely heavily on observations of filtration performance from noisy sensor data to assess membrane fouling conditions and lab-based protocols for plant maintenance, often leading to inaccurate estimations of future performance and delayed membrane cleaning. This challenge is further compounded by the difficulty in integrating existing complex mechanistic models with the Internet of Things (IoT) systems of wastewater treatment plants (WWTPs). By harnessing data obtained from WWTPs, along with innovative data denoising and model training strategies, we developed a machine learning application (MBR-Net) that is capable of forecasting membrane fouling, as indicated by permeability, for a full-scale submerged MBR plant in real time. We show that the trained model can effectively predict one-day-ahead changes in irreversible fouling under different desired fluxes, cleaning conditions and feedwater conditions (with MAPE < 6.45%, MAE < 3.71 LMH bar^–1, and R² > 0.87 on two independent testing sets). Although data availability presented certain limitations in the model development process, the current results demonstrate the significant value of machine learning in membrane fouling predictions and in providing decision support for fouling mitigation strategies in full-scale WWTPs.