An anomaly is not a diagnosis
Unusual flow or pressure can indicate leakage, a valve change, a new customer, instrument drift or a telemetry problem. Treat a detection model as a way to prioritise investigation. Separate detection, localisation and confirmation.
Combine complementary evidence
- Compare measured district inflow with an expected demand profile.
- Use minimum-night-flow analysis with an allowance for legitimate night use.
- Compare pressure residuals with a calibrated hydraulic model.
- Consult work orders, valve operations and known sensor faults before escalation.
Evaluate alerts in operational terms
Track precision, missed events, detection delay and localisation usefulness. A low average prediction error can still produce too many nuisance alerts. Benchmark under sensor outages and changing demand, then validate on independently confirmed events.
A safe path to use
Begin with retrospective analysis and then a shadow deployment in which staff review alerts without automatic control. Establish ownership of triage and feedback. A connected hydraulic model can add physical context, but it cannot compensate for unreliable network topology.
Distinguish an anomaly from a confirmed leak
A sudden flow increase can reflect a burst, an operational transfer, a meter issue or an unusual customer demand. A gradual night-flow rise may indicate leakage but can also reflect changing legitimate use or zone boundaries. An anomaly detector identifies a deviation from its learned or configured baseline; it does not directly identify the physical cause.
Combine the signal with pressure, valve and pump states, customer events and data-quality evidence. Preserve the distinction between a model alert, an investigated event and a confirmed repair. These are different labels and support different performance claims.
Build the monitoring boundary correctly
For a district or supply zone, establish all inflows, outflows, storage changes and the state of boundary valves. An unmeasured open interconnection can look like unexplained demand. Align timestamps and account for meter resets, gaps and drift before calculating a residual.
Night-flow methods require a defensible estimate of legitimate night use. A commercial or industrial zone may not follow residential assumptions. Changing pressure can also change leakage, so a comparison across operating states should account for the hydraulic context.
Evaluate event detection at an actionable threshold
Suppose a hypothetical method generates 20 event alerts, of which 8 correspond to confirmed leaks. There were 10 confirmed leaks in the evaluated period. On the stated event-matching basis, precision is 8/20 = 40% and recall is 8/10 = 80%. These values do not say whether the alerts arrived early enough or located the event accurately.
Define how multiple alerts for the same incident are grouped, the allowable matching window and what counts as confirmation. Otherwise, repeated alerts can inflate apparent performance. Also report time to detection, location usefulness, false alerts per operating period and the investigation workload.
If uninvestigated alerts are automatically labelled false, the test may penalise real but unconfirmed leaks. If only detected leaks are recorded in the reference set, recall may be overstated. Describe how the event register was established and its known gaps.
Use a layered investigation workflow
- Check whether the sensor and data feed are behaving normally.
- Review known pump, valve, transfer and customer events over the same interval.
- Compare the signal with neighbouring pressure and flow observations.
- Rank the remaining event for investigation using consequence and operational access.
- Record the field finding and update the event register without overwriting the original alert.
A machine-learning score can support the ranking, while hydraulic sensitivity or additional sensors can help narrow the location. Do not claim a precise leak coordinate when the available measurements only support a broad zone-level anomaly.
Pilot for operational usefulness
Use separate periods or events for development and evaluation, retaining the operational conditions that produced false alerts. Start in shadow mode and review how an engineer would have acted. Compare with existing night-flow, threshold or rule-based approaches under the same workload.
Track confirmed outcomes and data drift after deployment. A boundary change, new major customer or sensor replacement can invalidate the previous baseline. The data-quality workflow and zone-boundary guide provide the engineering foundation for interpreting a detection result.
Sources & further reading
- EPANET: water distribution modelling ↗US Environmental Protection Agency · 2020 release; living resource
External source · Checked 24 September 2026 - AI Risk Management Framework ↗National Institute of Standards and Technology · 2023 framework; 2024 generative AI profile
External source · Checked 24 September 2026
Source findings are distinguished from editorial interpretation. Apply current local criteria and project evidence when making engineering decisions.