An Uncertainty-Aware Leakage Detection Framework Based on Monte Carlo Dropout Neural Network

Leakage is a critical challenge in water distribution systems, leading to significant water and economic losses. Traditional detection methods are either invasive (costly and impractical for large networks) or non-invasive, including hardware-based, model-based, and data-driven approaches. While hardware solutions are expensive and model-based approaches are computationally intensive, data-driven methods offer a more scalable alternative. However, their deterministic predictions provide no quantification of uncertainty, limiting their reliability for real-time decision-making. To address this gap, we propose a novel, non-invasive leakage detection framework based on a Monte Carlo Dropout (MCD) neural network. An MCD model integrates Monte Carlo sampling with dropout regularization, enabling both uncertainty quantification and overfitting prevention. By using predictive distributions rather than single outputs, we quantify epistemic uncertainties and establish robust detection thresholds, ensuring that leakage detection explicitly incorporates confidence levels in the model’s predictions. We evaluate the framework on the Net3 benchmark network using EPANET simulations to generate normal and leakage events. The MCD model is trained solely on leak-free data, enabling it to detect anomalies when leakage occurs during inference. Results show that incorporating uncertainty intervals ensures detection reliability, reduces false alarms, and supports risk-aware operational decisions. This framework offers a scalable, real-time, software-based solution for uncertainty-informed leakage detection in modern water utilities.