Work

Process-Informed Latent-State Discovery and Interpretable Ensemble Learning for Full-Scale Wastewater Treatment Monitoring

Year: 2026
Type: article
Abstract: Reliable wastewater treatment monitoring requires methods that convert multivariate sensor streams into operationally meaningful states. This study proposes a process-informed, cluster-guided framework integrating unsupervised regime discovery, semantic state labeling, and interpretable tree-based classification. Online hydraulic, aeration, thermal, dissolved-oxygen, nitrate, and ammonium data from the full-scale Tilburg wastewater treatment plant were analyzed. Clustering identified four recurrent conditions: low-load operation, stable nitrification, oxygen-limited operation, and shock-load/transition conditions. The derived labels were predicted from routinely available online inputs using Decision Tree, Random Forest, and XGBoost classifiers, which achieved accuracies of 0.960, 0.988, and 0.989, respectively. Misclassifications were concentrated near boundaries between neighboring states, indicating gradual process transitions rather than widespread model instability. Model interpretation linked predictions to physically meaningful drivers, particularly influent loading, staged aeration, and temperature. The framework therefore provides an accurate, interpretable basis for real-time monitoring, early warning, operational decision support, and future integration with plant control systems in practical applications. Keywords: latent operating regimes; semantic cluster labeling; full-scale WWTP analytics; nitrogen-conversion dynamics; aeration-control intelligence; explainable tree ensembles; weakly labeled process data
Language: English
Subfield: Pollution
Open Access status: diamond
OpenAlex ID: W7235492930
FWCI:
Cites: 50
Cited by:
Related to: 10