Demonstration · Water networksSolving live in your browser

Find the leak. Keep the doubt.

The district meter shows night flow up by a fifth. Somewhere under these streets a main has split. Pressure loggers and a hydraulic model narrow the search, and the model says where the next logger will help most. Click any junction to move a logger and watch the probability map change.

District: 70 junctions · 100 pipes
Mains: 9.7 km
Supply: PRV at 72 m head
Hypotheses: 840 hydraulic solves
Loggers: up to 8

03:00 24:00

Loggers record pressure through the night; the model compares 02:00–04:00 with the previous night.

Detect, then locate.

Leaks are found at night, when demand is low and the extra flow stands out. The district inlet meter catches the rise. Pressure tells you where it is, because water heading for the leak loses head along the way, and the closer a logger is, the bigger its drop. Every logger drops by tens of centimetres; the location lives in the few centimetres of difference between them, so the answer is a probability map, not a pin.

Seven nights at the district inletL/s · 15-minute readings

Inlet flowNight flow before the burst

How much main to surveychance of finding the leak against length walked

Readings against the modelpressure change from the previous night, cm

Leak sizeposterior over 12 sizes, L/s at 40 m

What runs underneath.

Hydraulics

Hazen–Williams pipes solved with the Global Gradient Algorithm, the method inside EPANET: Newton steps on flows and heads with a Cholesky solve of the node matrix. The leak is a pressure-dependent emitter, so it leaks more where pressure is higher.

h = 10.67 L Q^1.852 / (C^1.852 D^4.87) q_leak = C_e √p

Hypotheses

Every junction is tried with twelve leak sizes from 0.8 to 10 L/s: 840 full network solves at 03:00, warm-started from the leak-free state. Each gives the pressure change at every junction and the extra inlet flow.

Inference

P(j,s | y) ∝ Π N(y_k; Δp_k(j,s), σ²) × N(ΔQ; ΔQ(j,s), σ_Q²)

σ combines the logger's accuracy with 2 cm of model error. The flow meter is trusted to ±0.6 L/s because legitimate night use varies.

Next logger

For each junction without a logger, the model simulates what that logger would read under hypotheses drawn from the current posterior, and how much the location uncertainty would fall. The largest expected drop in entropy, in bits, wins.

What this model leaves out

The district is invented. Demands are known up to their daily pattern, pipe roughness is known, and there is one leak at a junction rather than anywhere along a pipe. Real deployments calibrate roughness and demands from logger and meter history, use acoustic loggers alongside pressure, and treat several simultaneous leaks.

Shrink your search area.

We build hydraulic models, leak localisation and sensor-placement studies for water utilities, calibrated on your network data and logger history.

Discuss your project