Machine learning methodologies to predict possible water quality anomalies as a support tool for online monitoring of organic parameters

Abstract

[EN] Water Distribution Systems (WDSs) function to deliver high-quality water in major quantities. While standard water quality parameters are monitored at waterworks, it is still a challenge to monitor water quality in the WDS network itself. While mostly hydraulic parameters are frequently monitored and modelled in drinking water networks in Germany, the measurements of specific organic and bacteriological water quality parameter are still done offline which can take hours or even days which might be too late to react to possible water events. This study utilizes water quality data of a Utility in Hamburg, Germany to train machine learning algorithms to predict possible anomalies in specific water quality parameters which can indicate the necessity for more thorough investigations. While a large amount of water parameters is utilized and checked for deviations from the normal distribution, the input features to train the machine learning algorithms are just parameters which can be measured online like pH, temperature, total cell count of bacteria and the organic content of the water sample. A parallel study uses innovative online testing methods like fluorescence spectroscopy and flow cytometry in batch and flow experiments with the overarching goal of validating the trained algorithm to develop a wholesome online monitoring and warning system for drinking water anomalies. Various algorithms like Random Forest, Gradient Boosting, Decision Tree, Logistic Regression and Artificial Neural Networks are trained to predict whether the water samples indicate possible water quality anomalies. First results of this study show promising possibilities for a data driven online water quality prediction methodology which can help to digitalize the water sector immensely.