EU AI Act transparency duties apply now; high-risk duties from December 2027. Check your exposure
Insights About us Careers
Contact us
Energy & Utilities

AI for Water and Wastewater Utilities

Water utilities are judged publicly on leakage and pollution, both of which are largely invisible without analytics, and both of which now carry regulatory and reputational consequences that dwarf the cost of measuring them properly.

Tier 2
Our depth here
Leakage
The public number
Buried assets
Invisible by nature

Almost everything valuable in a water utility depends on knowing what is happening in pipes nobody can see, which makes the sector unusually dependent on inference from indirect measurement.

In one paragraph

AI for water and wastewater utilities covers leakage detection and localisation, treatment process optimisation, sewer network and storm overflow prediction, buried asset deterioration modelling, water quality monitoring, and demand forecasting.

Leakage, and what the network data can actually tell you

Leakage is the number a water utility is judged on publicly, and finding leaks in buried pipes is an inference problem. District metered areas, pressure sensors and acoustic loggers give indirect evidence, and how much they give depends on how the network was instrumented.

  • Minimum night flow remains the workhorse. Consumption at the quietest hour is the clearest leakage signal, and anomaly detection on it works well.
  • District boundaries must actually be closed. A district metered area with unrecorded open valves measures nothing reliably, and this is extremely common.
  • Pressure transients carry burst signatures. High-frequency pressure data localises events, and it needs sampling rates most telemetry does not provide.
  • Acoustic loggers plus analytics beat either alone. Correlation across loggers narrows the search area that crews then walk.
  • The output is a search area, not a location. Narrowing a crew's search from a kilometre to fifty metres is the realistic and valuable result.
Worth knowing

Check that your district metered areas are actually sealed

A large share of leakage analytics disappoints because the district boundaries leak in the accounting sense: valves open that records show closed, cross-connections nobody documented, meters that under-register at low flow. The analysis assumes a closed system and the system is not closed. Verifying boundary integrity is unglamorous field and data work, it usually precedes any modelling, and skipping it produces confident findings about districts that do not exist as measured.

Storm overflows and the accountability problem

Combined sewer overflow events have moved from an operational detail to a matter of public and regulatory scrutiny, with monitoring obligations and published data. That changes what analytics is for: not only reducing spills but evidencing what happened and why.

Rainfall-driven prediction is tractable

Sewer response to rainfall is physically driven and reasonably predictable given catchment characteristics, antecedent conditions and forecast intensity. Predicting which assets will spill under a forecast enables intervention where interventions exist.

Blockage and infiltration are distinguishable

A spill during dry weather indicates blockage or infiltration rather than hydraulic capacity, and the two have different signatures in level and flow data. Separating them directs entirely different work.

Evidence quality is now part of the deliverable

Where events are published and scrutinised, the defensibility of the data matters as much as the analysis. Monitor reliability, gap handling and event classification rules should be documented rather than implicit.

Prediction without intervention capacity is reporting

Knowing a spill will occur is only valuable if something can be done — storage, pumping, network control, or a crew clearing a blockage before rain arrives. We ask what the intervention is before building the prediction.

Treatment optimisation and buried asset deterioration

ApplicationValueThe catch
Chemical dosing optimisationHighCompliance limits are hard constraints, never trade-offs
Energy optimisation on pumpingHighLarge cost, quality-neutral, easiest to approve
Aeration control in wastewaterHighOften the largest single energy consumer on site
Water quality event detectionModerateSensor coverage and reliability decide what is possible
Mains deterioration modellingGoodFailure history exists here, unlike most energy assets
Sewer condition from CCTVGoodLarge survey image volumes; grading consistency is the win
Worth knowing

Water is one of the few utilities with real failure data

Mains burst frequently enough that deterioration modelling has something to learn from — unlike transformers, turbines or major plant. Age, material, diameter, soil, pressure regime and burst history support genuine failure rate modelling by cohort, which makes replacement prioritisation considerably better founded here than in electricity. It is worth exploiting, because it is unusual in the sector.

Process

How an engagement runs

Boundary and sensor integrity are verified before any analysis depends on them.

Weeks 1 to 3

Network and data integrity

District boundary verification, telemetry resolution and reliability, and whether the data supports the intended inference.

Weeks 4 to 9

Data foundation

Flow, pressure, rainfall, asset and quality data joined, with sensor reliability characterised.

Weeks 10 to 15

Build

Leakage detection and localisation, overflow prediction, or deterioration modelling by cohort.

Weeks 16 to 19

Field validation

Findings verified by crews, with search-area accuracy and false positive cost measured.

Ongoing

Operation

Continuous monitoring, retraining as the network changes, evidence quality maintained for published reporting.

Deliverables

What you receive

Smaller search areas for crews, fewer spills, and evidence that stands up publicly.

01

Network integrity assessment

District boundary verification and sensor reliability, before anything depends on them.

02

Leakage detection and localisation

Minimum night flow anomaly detection with pressure and acoustic correlation.

03

Overflow prediction

Rainfall-driven, with dry-weather blockage separated from hydraulic capacity events.

04

Deterioration modelling

Failure rates by cohort, exploiting the failure history water actually has.

05

Treatment optimisation

Dosing and energy, with compliance limits encoded as hard constraints.

06

Defensible evidence

Documented monitor reliability, gap handling and event classification for published data.

Fit check

Is this the right starting point?

Worth being direct. There are situations in water and wastewater where custom AI work is the wrong spend, and those are listed rather than buried.

Worth doing if

  • Leakage is a regulatory target and detection relies on manual survey.
  • District metered areas exist but boundary integrity has never been verified.
  • Storm overflow events are published and you cannot predict or explain them well.
  • Mains replacement is prioritised by age rather than by modelled failure rate.
  • Pumping or aeration energy is a major cost that has never been optimised.

Do something else if

  • Telemetry resolution is too coarse for the intended inference and cannot be improved.
  • There is no crew capacity to investigate findings, so nothing gets verified.
  • You want prediction where no intervention exists to act on it.
  • Compliance limits are being treated as parameters to optimise against.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

Why does our leakage analytics underperform?

Very often because the district metered areas are not actually sealed — valves open that records show closed, undocumented cross-connections, meters under-registering at low flow. The analysis assumes a closed system and the system is not closed, so it produces confident findings about districts that do not exist as measured. Verifying boundary integrity is unglamorous field and data work and it usually has to come first.

How precisely can AI locate a leak?

To a search area rather than a point, and that is the useful result. Narrowing a crew's search from a kilometre of main to fifty metres changes the economics of leak detection substantially. Pressure transient data and correlation across acoustic loggers do most of the work, and both depend on sampling rates that standard telemetry frequently does not provide — check that before scoping.

Can we predict storm overflow spills?

Reasonably well, because sewer response to rainfall is physically driven and depends on catchment characteristics, antecedent wetness and forecast intensity. The question we ask first is what the intervention would be — storage, pumping, network control, or clearing a blockage before rain arrives. Prediction without an available intervention is reporting, and it is worth establishing which you are buying.

Is deterioration modelling realistic for buried assets?

In water, yes, and that is unusual in this sector. Mains burst frequently enough that failure rate modelling by cohort — age, material, diameter, soil conditions, pressure regime — has real data to learn from, unlike transformers or turbines. It makes replacement prioritisation considerably better founded here than in electricity, and it is worth exploiting.

Can AI control chemical dosing?

It can recommend dosing, and compliance limits must be hard constraints rather than objectives to trade against. Drinking water and discharge standards are legal limits with public health consequences, so they come from your compliance function as inviolable inputs. Advisory dosing optimisation to an operator is a considerably shorter path to value than anything closing the loop, and it captures most of the benefit.

Tell us what the problem looks like.

Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.