Sizing a pump, a main or a storage tank is the part of water system design that is well understood. The equations have been stable for decades, the software is mature, and any competent engineer will land in roughly the same place given the same inputs.
That last clause is where projects go wrong. In our experience the design is rarely the weak link. The inputs are.
Why Designs Fail in the Field
A network model is an argument about the future built from records of the past. When those records are thin, the design inherits the gap and nobody finds out until commissioning:
- Demand is assumed, not measured. Per capita figures from a design manual are applied to a population estimate, and the resulting diurnal curve has never been checked against a meter.
- Asset records disagree with the ground. Diameters, materials and valve states on the drawing differ from what was actually installed, sometimes by decades.
- Monitoring data has holes. SCADA and logger series lose stretches to power cuts and instrument faults, and a straight line drawn through the gap quietly becomes a calibration target.
- Peak factors carry the uncertainty. When the underlying data is weak, safety factors absorb the doubt. That is oversizing, and it is paid for in capital cost and in pumps running badly off their best efficiency point.
Where the Physics Belongs
Hydraulics is not a place for learned models, and we do not treat it as one. Pressure, headloss, fire flow, surge and pump curves follow physical laws, and a network you are about to change has no operating history to learn from.
A new main, a new zone boundary or a new pumping regime is a stress the system has never been put under. That is exactly the case where a model trained on past behaviour has nothing useful to say and a solver does. We build these in EPANET, and the same principle governs our work in water distribution modeling, stormwater and surface water.
Where Machine Learning Earns Its Place
Machine learning is useful here, but on the data feeding the model rather than inside the hydraulics. Four jobs in particular:
Demand Forecasting
District demand a day to three days ahead, learned from consumption, weather and calendar patterns. This is an input a hydraulic solver requires and cannot generate for itself, and it is a far better basis for an operating study than a textbook diurnal curve.
Filling Gaps in the Record
Logger and SCADA series are reconstructed across their gaps before calibration rather than after. This matters more than it sounds. Calibrating against an interpolated gap does not produce a slightly worse model, it produces a confidently wrong one, because the fitted roughness absorbs the error.
Anomaly Detection on Night Flow
Minimum night flow is where leakage separates most cleanly from legitimate consumption. A model trained on each district's own history flags a departure from its established baseline, which catches a developing burst that no fixed threshold would.
Separating Instrument Faults From Real Behaviour
A drifting sensor and a genuine hydraulic event look similar in raw data. Screening each instrument against its own history keeps a fouled meter from being calibrated into the design as though it were truth.
How to Judge a Design Before You Build It
If you are commissioning this work rather than doing it, these questions separate a model that will hold from one that will not:
- What was it calibrated against, and how well? Ask for the field pressures and flows used, and the residual error. A model that has never reproduced a measured condition is an opinion with a mesh.
- Where did the demand figures come from? Measured, or taken from a manual. If measured, over what period and how many meters.
- What happens to the gaps? Ask specifically how missing data was handled. Silence here usually means straight line interpolation.
- What decision does this design close out? A model with no decision attached to it is a deliverable, not an asset.
The Point of the Exercise
Good mechanical design is necessary and it is not sufficient. The value sits in whether the system you build behaves the way the model said it would, and that is decided by the quality of the record you calibrated against long before anything is specified.
We build the monitoring into the work for that reason rather than treating it as a later phase. If you are scoping a network study or a design review, our water consulting and modeling services set out how we approach it.