Machine Learning Applied to Chlorine Residual Data
Author(s)
Helm, Wiley Speir
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Chlorine disinfection systems are ubiquitous in drinking water treatment plants and provide a method for reliable, safe, potable water. The concentration of free chlorine residual (FCR) leaving drinking water treatment plants must be between a federally regulated margin of 4.00 and 0.20 mg/L as Cl2. Dosing chlorine is tricky as oxidizable constituents in the water will consume the FRC generating a chlorine demand and altering the final concentration. The chlorine dose is often determined by the plant operator who relies on personal experience to predict the chlorine demand and the FCR leaving the plant. A machine learning model was applied to historic water treatment plant data to see if the FCR could be predicted with readily available water quality data. The model used a CatBoost regression tool, based on a form of gradient boosting, to predict the FCR. The data spanned from March 2021 to March 2022, and a data frame of 33 input parameters and 7,851 rows of data was used to train the model. The model went through four iterations, with the best predictive results providing a coefficient of determination, R2, of 0.937. Shapely additive explanation values were used to interpret the results of the model. This tool provided elegant visual interpretations into the function of the model and how it used individual parameters to make predictions. It was found that the model relied heavily on non-causal parameters to make the most accurate predictions. The relationship, if any, between the non-causal parameters and the FCR is not clear.
Sponsor
Date
2022-12-13
Extent
Resource Type
Text
Resource Subtype
Thesis