An Interpretable Stacked Ensemble Model for Classifying Abnormal Non-Revenue Water

DOI

https://www.doi.org/10.59256/indjcst.20260503048

Published In

Volume 5, Issue 3 (2026)

Published Online

September-December 2026

Pages

417-423

Article Type

Review Article

Check for updates

In developing countries, non-revenue water (NRW) is a significant operational and financial challenge, but is generally managed in a reactive way. This paper presents and critically assesses an interpretable, imbalance-aware stacked ensemble for abnormal NRW detection at District Metered Area (DMA) level, leveraging 43 months (January 2023–July 2026) of operational and commercial data from Embu Water and Sanitation Company (EWASCO), Kenya. The analytical data set included 742 DMA-month observations from 24 zones, of which those with NRW values over the 20% sector value were considered Abnormal. A class-weighted Logistic Regression was used as a meta-learner to combine the Random Forest, Support Vector Machine, and XGBoost base learners, while the volumetric terms were not included in the base learners as this would lead to target leakage. Imbalanced metrics were used to evaluate performance, and SHAP and permutation importance were employed for providing interpretation. Average NRW was 45.1%, zone averages ranged between approximately 25% to 79%, and none of the zones achieved the NRW average. The target was very much unbalanced in the other direction than was assumed (94.1% Abnormal). The ensemble's performance on the held-out test set was accuracy 0.853 (trivial majority classifier 0.933), precision 0.984, recall 0.857, F1 0.916, PR-AUC 0.980 (no-skill baseline 0.933), and balanced accuracy 0.829 (0.500), but it failed to outperform its base learners. The variables that had the greatest influence were lagged system input volume and lagged system billed consumption (which was not an early warning indicator, but rather hydraulic persistence).