Forecasting bacteriological presence in treated drinking water using machine learning

A novel data-driven model for the prediction of bacteriological presence, in the form of total cell counts, in treated water exiting drinking water treatment plants is presented. The model was developed and validated using a year of hourly online flow cytometer data from an operational drinking water treatment plant. Various machine learning methods are compared (random forest, support vector machines, k-Nearest Neighbors, Feed-forward Artificial Neural Network, Long Short Term Memory and RusBoost) and different variables selection approaches are used to improve the model's accuracy. Results indicate that the model could accurately predict total cell counts 12 h ahead for both regression and classification-based forecasts—NSE = 0.96 for the best regression model, using the K-Nearest Neighbors algorithm, and Accuracy = 89.33% for the best classification model, using the combined random forest, K-neighbors and RusBoost algorithms. This forecasting horizon is sufficient to enable proactive operational interventions to improve the treatment processes, thereby helping to ensure safe drinking water.