The water-leaks programme studies leak detection and localisation on BattLeDIM's L-Town, a simulated water network with two years of sensor data and a published leak schedule. Nothing here is a field result.
The last step fitted each recently dated leak source with a step or a ramp. The labels came out right: every slow leak's nearby source was a ramp, and ramps are rare among the candidates (about 1 in 6). But it then timed each change against the alarm, and a slow leak that is still rising looks days older than it is, so the newest source won again.
The rule
Use the label, drop the timing. If any source dated in the last 30 days fits as a ramp, pick the ramp rising fastest. If none does, pick the newest source. The rule and five predictions went into the research log before it was run.
What happened
| rule | within 300 m (of 32) | bursts (15) | slow leaks (17) |
|---|---|---|---|
| steepest ramp, else newest (this step) | 15 | 9 | 6 |
| fastest-rising source (step 22) | 14 | 8 | 6 |
| newest source (step 21) | 13 | 10 | 3 |
| regression residual, no tracker (step 11) | 15 | 10 | 5 |
| one-leak oracle (upper bound) | 28 | 13 | 15 |
Four of the five predictions held: more than 14 overall, at least 6 slow leaks, at least 9 bursts, and where a ramp source sits within 300 m of the leak, the rule lands within 300 m for 11 of 13. It also placed two slow leaks that no earlier tracker rule found. The fifth prediction, a significant win over the newest-source rule, failed (3 leaks against 1, P = 0.63), as the plan expected.
Why this is not a method claim yet
The rule was chosen after seeing the last step's per-leak shapes, on the same 32 leaks, so 15 is an in-sample number. It also only ties the plain regression residual's 15, and it runs on a residual that removes demand with the generator's own model, which a utility does not have. Fifteen leaks have no ramp-shaped source at all; there the rule is just "the newest", right for 4.
The next test is out of sample: the same rule, frozen, on simulated leak years from the L-Town model, where the leaks are new and a paired test has enough of them to mean something.
Code, results and the per-leak table: experiments/step24-shape-chosen/.