Data Driven Approaches to Enhance Choke Valve Reliability

Wellhead choke valves take the largest pressure drop in the production system, and when they erode the result can be a loss of containment. The authors tested common velocity rules of thumb against five years of choke inspection records and found that the usual limit did not separate failed valves from healthy ones, while a lower threshold did. They then showed that subsurface properties known before a well is produced predict choke damage better than flow velocity. The result is a data-driven basis for choke selection, alerting and refurbishment planning.

Key results

Of 133 Cv-alert inspections, 48 found damage and 85 did not, so the existing tool was about 36.1% accurate
The 25 m/s velocity limit barely separated the groups: 41.2% of valves above it were damaged versus 36.9% below (p = 0.74)
A 15 m/s threshold was the most significant (p = 0.00685), with failure ratios of 48.3% versus 21.95%. Adding it to the Cv alert would raise prediction accuracy from 36.9% to 48.3%
XGBoost was the top classifier, and subsurface features such as true vertical depth and porosity outranked velocity as predictors of valve damage

Approach

Combined 5 years of choke removal and inspection records with well-test data (rates, pressures, temperatures, choke size) and drilling, completion and perforation data for 133 wells
Reviewed ISA, API 14E and NORSOK guidance and the manufacturer's cage-velocity rule, converting the latter to an equivalent trim-exit velocity
Swept the velocity threshold from 1 to 50 m/s and ran Student's t-tests to find the most statistically meaningful cut-off
Summarised the time-series data into per-well statistics and trained more than 40 binary classifiers with 6-fold cross-validation and a hold-out fold, scored by AUC

The problem

There is no industry standard for choke-valve velocity limits, and the existing valve-coefficient (Cv) deviation alert produced many unnecessary valve pulls.

Tools & methods

Student's t-testXGBoost classifierAutoML model leaderboard (40+ models)cross-validation / AUCCv monitoring modelISA / API 14E / NORSOK velocity criteria

Summary written for this site from the published paper. For figures, data and the full method, read the paper on OnePetro.