A Machine Learning-Based Smart SCADA Framework for Predictive Fault Detection and Optimisation-Based Control of an Urban Wastewater Treatment Plant

Abstract

Supervisory control and data acquisition (SCADA) systems in urban wastewater treatment plants (WWTPs) traditionally act as passive data historians rather than predictive decision-support tools, leaving operators to react to effluent-quality excursions only after laboratory analysis confirms them. This study develops and fully implements, on a real, openly available case study [1–3], an end-to-end machine learning-based smart-SCADA architecture that (i) discovers and classifies plant operational states, (ii) predicts effluent quality ahead of laboratory confirmation using leakage-free soft sensors, and (iii) closes the loop with an event-triggered, optimisation-based controller. The case study is the UCI Water Treatment Plant dataset (527 daily records, 38 operational variables, expert-labelled operational states) from an urban plant in the Barcelona metropolitan area. Unsupervised clustering of the standardised feature space showed only weak agreement with the plant experts' own fault categorisation (adjusted Rand index = 0.062), motivating a supervised approach: a class-balanced support-vector classifier achieved the best fault-detection performance (ROC-AUC = 0.999, balanced accuracy = 0.928), while an unsupervised Isolation Forest achieved comparable discrimination without requiring any labelled fault history (ROC-AUC = 0.999). Soft-sensor regression models predicting effluent BOD5, COD and suspended solids from upstream measurements and a one-day autoregressive term achieved hold-out R² of 0.297–0.301. Coupling these models to a derivative-free, surrogate-model-based controller that searches within the historically trusted operating envelope reduced the total predicted exceedance of illustrative regulatory discharge limits by 93.8% on a 105-day chronological hold-out, while fully resolving 28.6% of individual limit exceedances, using an intervention triggered on only 10.5% of days. The results, reported with explicit uncertainty and limitation analysis rather than optimistic point estimates alone, indicate that a modest, software-only augmentation of legacy SCADA infrastructure can meaningfully improve predictive fault awareness and compliance margins without new hardware instrumentation.