The numbers, and their limits.
Tournament t001, 1500 held-out simulated wells (deepwater and gas-lifted tiebacks; a shallow satellite held out as unseen); the truth is a six-state riser model of a different structure from the estimator's. Plus 315 real instances from a public well-event dataset, tree gauge hidden. The bar and the result are on the proof page.
Tree-pressure error on the hidden gauge
See the tables below.
Interval coverage
See the tables below.
Slug lead time
See the tables below.
False alarms per day
See the tables below.
Same cases, same rules
Scored against, untouched
- Oscillation threshold on topside pressure
- Published event classifiers
- ARX and NARX models
- Gradient-boosted trees
- Recurrent and convolutional soft sensors
Thresholds set on validation data and frozen for every test, including the cases the engine never saw. The truth model differs from the engine's own.
The bar, set before the first run
Match black-box soft sensors on unseen wells, with calibrated intervals.
Simulated split, wells of a geometry never seen in training: tree-pressure error 4.3 bar for the observer bank against 39.2 bar for the best supervised soft sensor (ARX soft sensor (trained on the tree gauge)); difference 90 % interval -35.1 to -34.6 bar; bank coverage 47 %. Real wells (a public well-event dataset, 315 instances on 9 wells, every well unseen): 37.8 bar for the bank against 21.9 bar for the best soft sensor trained on the other wells' gauge (Recurrent soft sensor (trained on the gauge)); difference 90 % interval 7.0 to 26.6 bar, wells resampled; bank coverage 39 %.
First proof
Three to four weeks: a public labelled well-event dataset, real instances only; tree pressure estimated from topside channels, hidden and scored on held-out windows and unseen wells; against the black-box soft sensors.
Tree pressure, never measured
| Error, seen geometry | Error, unseen geometry | Inside band, seen | Inside band, unseen | |
|---|---|---|---|---|
| Tieback: observer bank (never sees the tree gauge) | 8.9 bar | 4.3 bar | 81 % | 47 % |
| Recurrent soft sensor (trained on the gauge) | 2.8 bar | 45.2 bar | 92 % | 0 % |
| Gradient-boosted soft sensor (trained on the gauge) | 2.6 bar | 46.5 bar | 90 % | 0 % |
| ARX soft sensor (trained on the tree gauge) | 8.2 bar | 39.2 bar | 88 % | 0 % |
Median per-well RMSE over the four-hour scoring window. The soft sensors were trained on the tree gauge of 1,000 training wells; the observer bank never sees it. Unseen geometry: the shallow satellite family, never in any training.
Slugs, at 0.4 false alarms per well-day
| Slugs caught | Lead after onset | False alarms / day | |
|---|---|---|---|
| Tieback: observer bank (never sees the tree gauge) | 25 % | 43.9 min | 0.3 |
| Event classifier (published design) | 20 % | 0.8 min | 0.4 |
| Oscillation threshold | 4 % | 12.6 min | 0.4 |
Thresholds frozen on validation. Onset: the first cycle in which the low point blocks the gas (from the truth).
Real wells: tree pressure against the hidden gauge
| Median error | Median bias | Inside band | |
|---|---|---|---|
| Tieback: observer bank (never sees the tree gauge) | 37.8 bar | -28.2 bar | 39 % |
| Recurrent soft sensor (trained on the gauge) | 21.9 bar | 13.8 bar | 53 % |
| Gradient-boosted soft sensor (trained on the gauge) | 33.1 bar | -2.1 bar | 0 % |
| ARX soft sensor (trained on the tree gauge) | 33.8 bar | 1.3 bar | 44 % |
A public labelled well-event dataset, real instances only: 308 normal, 7 hydrate on 9 wells; first hour (labelled normal) to identify, next two hours scored. Excluded and counted: 28 shorter than 3 h; 262 tree or topside pressure missing; 3 implausible pressures; 32 severe slugging: no labelled-normal start; 343 flow instability: no labelled-normal start. Well geometry is not in the data: every well gets the deepwater prior. Soft sensors trained leave-one-well-out on the other wells' gauge.
What is proven. What is not.
Proven, in simulation
- On an unseen well geometry, tree-pressure error 4.3 bar against 39–47 bar for soft sensors trained on the gauge of the other geometries (simulation, t001)
- Never reads the tree gauge; a guard raises if an estimator tries
- Catches 25 % of slugging wells at 0.28 false alarms a day; oscillation threshold 4 %, event classifier 20 % (simulation)
Not proven
- In simulation, on the unseen geometry the 90 % band covers 47 % of the truth
- On real wells: 315 instances on 9 wells of a public well-event dataset, 37.8 bar median error against 21.9 bar for a recurrent soft sensor trained on the other wells' gauge; band covers 39 %
- On seen geometries the soft sensors win: 2.6–2.8 bar against 8.9 bar
- Slug alarms come late: a median 44 minutes after onset
- Largest stable choke: 98 % median error on the main families
- No well identifiable from a quiet window: every band comes from priors
- No real slugging instance usable: none starts with a labelled-normal hour; no well geometry in the public data
The first real number comes from your data.
A month of your exports, replayed offline, gives the figure that matters on real operations. We will show it to you whether it is good or bad.