This morning's step 15 ended with a confident sentence. On the BattLeDIM L-Town benchmark, a perfect residual (the exact pressure effect of the leaks) helped a model-based leak localiser by only one leak, so we wrote that the hydraulic model must be what limits localisation. Step 16 tested that directly, and the sentence didn't survive.
The test
BattLeDIM publishes two network models. L-TOWN.inp is the nominal model the competitors got. L-TOWN_Real.inp is the model the organisers generated the data with. Every pipe differs between the two: diameters are a median 2.1% smaller in the real one, roughness is within about ±9%, and each node has its own demand pattern.
So we computed leak signatures from the real model itself. For each of the 32 leaks we restarted the real model at the hour we localise it, with the tank level and pump state the leak-free network had at that moment. Then we drew the leak's true flow at each of the 782 junctions in turn and recorded the 24-hour pressure change at the 33 sensors. That came to 25,056 EPANET runs, about 17 minutes on four cores. A check came first: the restarted leak-free runs match the continuous two-year simulation exactly.
With the true model, the true day, the true leak size and the exact residual, we expected at least 24 of 32 leaks within 300 m. We wrote that down beforehand, along with four other predictions.
The result
- the generator's own model, on each leak's day: 15 of 32 within 300 m;
- the nominal model from step 15: 16 of 32;
- naming the sensor whose pressure fell most: 14 of 32.
None of the differences comes near significance (paired sign tests, P = 1.0). All five predictions failed. The leak pipe's own end node scored best for only 5 of the 32 leaks.
Where the misses are
After the predictions were scored, we split the leaks by type:
- the 15 abrupt leaks: 11 within 300 m with the real model, 11 with the nominal model, 12 with the sensor;
- the 17 incipient (slow-growing) leaks: 4, 5 and 2.
We localise each leak at the first hour its flow reaches half its peak. For a slow leak that hour comes a median five weeks after onset, and the residual is measured against the day before onset. Over those five weeks the network's other leaks change too: by a median 17.8 m³/h, against the slow leak's own 6.7 m³/h. So most of the "leak's residual" is other leaks. Split at the median of that ratio, the real model places 12 of the 16 cleaner cases within 300 m and 3 of the 16 dirtier ones (Fisher exact P = 0.004). We chose this split after seeing the results, so it is a lead for the next step, not a finding.
What changes
For abrupt leaks with a clean residual, the nominal model already does as well as the exact one. On L-Town, the 2% diameter error doesn't cost localisation. The residual does: a baseline fixed when a slow leak starts goes stale while everything else in the network keeps moving. That is the part a utility could act on. A localiser for slow leaks needs a baseline that tracks the other changes in the network, not a better-calibrated pipe model.
Everything here is an oracle bound on a simulated network: the true model, leak times and sizes are things no utility has. The next step replays the other leaks through the generator's model, so each residual holds one leak only. That will separate contamination from what 33 sensors can actually resolve.
Code, the plan written before the run and the per-leak table are in research/water-leaks (step 16).