Interpretable Dynamic Prediction and Risk Assessment of Filtered-Water Turbidity in a Drinking Water Treatment Plant

Abstract

Filtered-water turbidity at the study plant was usually low during 2025, but several short excursions coincided with changes in raw-water and hydraulic conditions. The practical problem was not simply to predict those values: operators also needed to know which measurements were driving a warning and whether the predicted value represented a meaningful operational risk. We analysed 4,380 records from January to December 2025 and used observations from January to March 2026 as a separate follow-up set. The available signals covered raw-water turbidity, pH and flow, alum dose, river level, clear-water well level and filtered-water turbidity. Missing entries were handled according to their frequency, while extreme observations were retained and flagged because they could represent genuine plant events. Predictor relevance was examined with Pearson correlation, mutual information, LASSO and randossm-forest importance. XGBoost and LightGBM were then combined in a weighted model, with SHAP values used to inspect individual contributions. An ARX formulation was fitted separately to examine delayed responses, and entropy-weighted fuzzy grading was used to express risk. Clear-water well level, river level and lagged raw-water flow emerged as the leading variables. The ensemble obtained R2 = 0.834 on the training data, whereas the extended ARX model obtained R2 = 0.800. Of the 2025 observations, 90.6% fell in the safe class; every follow-up observation was also classified as safe. The framework therefore links a turbidity estimate to both an explanation and an operational risk label.