Date of Award

Spring 2025

Project Type

Thesis

Program or Major

Civil and Environmental Engineering

Degree Name

Master of Science

First Advisor

James Malley

Second Advisor

Fei Han

Third Advisor

Irina Zaikina

Abstract

Ultrafiltration (UF) membrane operation is strongly affected by fouling, where dissolved and suspended water quality constituents interact with each other and the membrane surface, leading to pore blockage. As a result, a fouled membrane loses filtration efficiency, reducing the volume of treated water produced and increasing operation and maintenance costs associated with membrane cleaning. At a Dutch full-scale facility in North Holland, fouling is indicated by both an increase in the amount of pressure required across the membrane, known as transmembrane pressure (TMP), and a decrease in permeability, represented by specific flux (SF). The facility performs hydraulic backwashing - both with and without chemical addition- occurs on a fixed frequency to scour readily removable foulants from the membrane surface. However, more persistent fouling that remains after backwashing can accumulate throughout filtration and must be removed with a more aggressive cleaning method. This is done through a cleaning in place (CIP) procedure, during which the membrane is soaked in two stages of chemicals for several hours to attack the fouling that remains after backwashing. The UF system consists of eight membrane blocks (i.e., groups of membranes), with six of them having viable data for modeling, resulting in six models tailored to each of the membrane blocks.The CIP is typically performed once TMP reaches 100 kPa. Currently, there is no early warning system to alert operators when this threshold has been reached, requiring frequent manual monitoring of TMP to determine an appropriate time to conduct a CIP. A previous intern of the utility investigated machine learning approaches to forecast TMP and found that a random forest (RF) model yielded the lowest prediction error. Building on that work, this study applies a regression-based RF approach to predict how SF will change in the next week, providing operators with an estimate of future SF with ample time (seven days) for CIP scheduling. The model uses current SF and SF from the previous week as inputs to predict the change in SF that will occur in the next week. Operational input data was collected from online UF data collectors, while water quality data was collected from both UF influent data collectors and from laboratory measurements. The resulting ‘Block Model’ for each of the six membrane blocks achieved R2 values exceeding 60% and mean square error (MSE) values below 3.20E-8. A combined ‘Universal Model’ achieved an R2 of 66%. Operators may consult this Universal Model to anticipate changes in SF and proactively plan CIP procedures.

Download

Share

COinS