This virtual event explored how machine learning augments traditional rules-based methods for customer-side water leakage detection. Rules work well for standard households (e.g., nighttime flow alerts) but struggle with non-standard usage and non-household properties, where intermittent, low-volume, or continuous processes obscure leaks.
Tackling Leakage through Machine Learning
Success depends on data quality, scalable engineering for millions of meters, and a closed feedback loop to validate outcomes, retrain, and maintain performance. While ML has higher compute costs, it unlocks savings through increased leak recovery, regulatory gains, and operational efficiency, with single-tool integration improving prioritization and reporting. Data privacy must be addressed through appropriate system design and governance.
Speakers
- Jason King, Business Development Manager, Arqiva
- Stephanie Edmonson, Senior Product Manager, Arqiva
- Jack Kilvington, Head of Data Science, Arqiva
Key takeaways
1. Rules Remain Reliable: Rules-based leakage detection remains effective for standard households (e.g., 2–4 a.m. no-flow rules) but struggles with non-standard usage such as industrial sites, high-use properties, and intermittent low-volume leaks.
2. Machine Learning Augments: Machine learning complements rules by learning patterns from data (supervised and unsupervised) to identify subtle and edge-case leaks that rules miss, enabling targeted, higher-value interventions.
3. Hybrid Strategy Wins: A hybrid approach is recommended: use rules to efficiently clear straightforward cases and apply ML to the 20% of meters that consume 80% of operational effort and often drive disproportionate leakage volumes.
4. Data-Driven Deployment: Successful ML deployment requires robust data quality, scalable engineering to handle millions of meters, and a closed feedback loop (verification, retraining, performance monitoring) to maintain real-world accuracy.
5. Cost-Saving Intelligence: While ML adds compute cost versus rules-based methods, it unlocks savings via earlier detection in complex scenarios, supporting regulatory performance and reducing water production, abstraction, and associated emissions.
