Application of machine learning (ML) techniques for water quality assessment in a water treatment plant

Abstract

Clean water is an important resource, one that is not readily available to everyone, especially in developing countries such as South Africa. This is due to a variety of reasons, including overpopulation, pollution, and lack of infrastructure in water treatment facilities. Gauteng Province, the economical hub, as well as the most populated province in South Africa receives its water from Rand Water, Africa’s biggest water treatment facility.

Rand Water, like many other water treatment facilities uses various processes to clean raw water. Throughout these procedures, sensors and communication devices monitor chemical parameters such as pH, alkalinity, chlorine etc. in real-time. This produces enormous amounts of complex data, known as big data. It has been reported that water treatment facilities use less than 40% of the data that they produce. This has undoubtedly deprived these facilities and authorities to fully exploit such data and draw some insights that can be actionable towards better water management and planning.

Machine learning, a subfield of artificial intelligence that allows software applications to accurately predict outcomes without being explicitly programmed, has presented some of the most useful tools for processing large datasets and has found application in medicine, finance, marketing, art, music, and environmental management among other areas.

This study aimed at applying machine learning techniques to assess water quality data at one of the treatment facilities at Rand Water. The approach involved: (1) cleaning and pre-processing datasets for nephelometric turbidity units (NTU) obtained from four filter houses; (2) developing algorithms that can effectively predict future turbidity levels, namely multilayer perceptron artificial neural networks (MLP-ANN), long-short term memory (LSTM), random forest (RF), and support vector regression (SVR); and (3) testing the robustness of the models under different model input and hyperparameter settings.

For the predictive analysis, two scenarios were developed: Scenario 1 were predictions using the complete datasets and Scenario 2 were predictions using datasets that were treated for outliers. The performances of the algorithms were monitored using three loss functions, namely: root mean square error (RMSE), mean squared error (MSE) and coefficient of determination (R2). The errors are expressed as absolute values.