Enhancing drinking water safety: Real-time prediction of trihalomethanes in a water distribution system using machine learning and multisensory technology

Prolonged exposure to high concentrations of trihalomethanes (THMs) may generate human health risks due to their carcinogenic and mutagenic properties. Therefore, monitoring THMs in drinking water distribution systems (DWDS) is essential. This study focused on the statistical modelling of THMs formation through multiple linear regression (MLR) method to develop simple predictive models that serve as preventive tools capable of alerting about potential increases in THMs within the water network. To achieve this, a dataset comprising 1192 observations of water quality measurements in the study area over five years was created. The independent variables selected to explain the formation of THMs were free residual chlorine (FRC), total organic carbon (TOC), conductivity, pH and turbidity. Then, following an exploratory analysis of the dataset using Pearson’s correlation matrix and an ANOVA test, multiple regression models were developed. In total, a total of two predictive models were built, based on data filtered by conductivity levels, with coefficients of determination (R²) of 0.64 and 0.47. The algorithms of these predictive models were integrated into the control center of the water company in the study area. On the other hand, a multisensory device was installed in a strategically located drinking water tank to measure the values of the independent variables used in the models. These measurements were transmitted online to the control center to continuously update the predictive models and provide real-time forecasts of THMs. Finally, model validation was performed by comparing the real-time predictions of the models with actual THMs levels obtained from laboratory analyses, achieving an average accuracy of 90 %.