A week-old baseline cannot find a slow leak: 12 of 32 on BattLeDIM, against an oracle's 28

On BattLeDIM L-Town, a residual a utility could compute (pressure against the same hours a week earlier) localises 12 of 32 leaks, and no lag does better.

Earlier today step 17 of our leak programme set a bar. On the BattLeDIM L-Town benchmark, if the pressure residual holds only one leak's effect, a correlation localiser with the uncalibrated nominal model puts 28 of 32 leaks within 300 m. Building that residual took an oracle, though: we replayed every other leak's published flow through the organisers' own model. A utility has no list of its other leaks.

So step 18 asked the obvious practical question. Can a residual with no leak list in it get anywhere near that bar? The simplest candidate is a moving baseline: take the day's mean pressure at each of the 33 sensors and subtract the same hours a week earlier. Old, steady leaks cancel, and so does the weekly demand pattern. We wrote down five predictions before computing a single lagged residual.

The result

residual (nominal model, q = 5 m³/h)within 300 mslow leaks (17)
pressure minus the same hours a week earlier12 of 324
the same, with demand change removed by the leak-free model11 of 323
regression residual (step 11)15 of 325
one-leak oracle (step 17)28 of 3215

Three predictions failed. The weekly difference does no better than the regression residual (paired sign test P = 0.45), and no better than just naming the sensor whose pressure fell most (12 of 32). Two held: the residual stays far below the oracle, and removing demand change makes no difference. The loss isn't demand.

Other lags fare no better. A one-day lag gets 12 (it keeps the abrupt leaks, 11 of 15, and loses nearly every slow one, 1 of 17). Three days gets 8, two weeks 11.

Why no window works

The published leak flows show what is in each window. Over a week, a slow leak's own flow grows by a median 2.3 m³/h. In the same week the network's other leaks change by a summed 2.7 m³/h. Only about a third of the slow leak's growth since onset is still inside the window. A short window drops the old leaks, but it also drops most of the slow leak. A long window keeps the slow leak, but by then the other leaks have changed as well. In 2019, 23 leaks overlap on L-Town, so no fixed window gives a slow leak to itself.

What it means

A cheap moving baseline won't close the gap between today's practice (about 15 of 32) and what 33 sensors can resolve (28 of 32). The other leaks have to be modelled rather than differenced away. One way is a sequential localiser that subtracts the modelled effect of leaks it has already found. Another is to fit several leak sources jointly to the residual. Either one has a clear bar now.

Caveats: the localisation hour is still an oracle (when the leak reaches half its peak flow), L-Town is a simulated network, and 32 leaks is a small sample. Code, the plan as written before the run, and the per-leak table are in research/water-leaks (step 18).

All writing · More in Research findings · All tools and research