Back to Projects
C
Data Science / Analysis

Choke Performance Prediction

ML Engineer (Research)

Scikit-learnRandom ForestGridSearchCVPandasPython

About the Project

A postgraduate research project benchmarking machine learning against classical empirical correlation for predicting oil production rate through surface chokes in mature Niger Delta wells. Built on 1,000+ daily observations from the Opukishi field spanning 7 wells (A–G) across 4 flow stations, the Random Forest regressor predicts oil rate in bbl/day from choke size, wellhead and tubing-head pressure, separator pressure, water cut, gas-oil ratio, temperature, viscosity, and oil specific gravity — then is scored head-to-head against the Gilbert correlation that petroleum engineers have used for decades.

What It's Made Of

Dataset assembly

1,000+ observations · 7 wells · 4 stations

Building one analysable dataset from daily production records across seven wells (A-G) and four flow stations in the Opukishi field, standardising units and column names between sources that did not agree with each other. Unglamorous and the precondition for everything else.

Field-condition cleaning

Shut-ins · workovers · sensor faults

A mature field spends a lot of time not producing. Shut-in periods, workovers, maintenance windows and sensor errors are flagged and removed deliberately, with each decision documented, because left in they teach the model that zero flow is normal and quietly destroy it.

Random Forest model

Scikit-learn · GridSearchCV

A Random Forest regressor predicting oil rate in bbl/day from choke size, wellhead and tubing-head pressure, separator pressure, water cut, gas-oil ratio, temperature, viscosity and oil specific gravity, tuned with GridSearchCV and examined through feature importance to see which production variables actually drive the prediction.

Gilbert benchmark & residual analysis

6 sequential notebooks

The part that convinced domain experts: scoring the model head-to-head against the classical Gilbert empirical correlation on the same held-out data, then analysing residuals across operating regimes. Gilbert showed systematic error in specific regimes that the Random Forest captured — an argument a single accuracy number could not have made.

Key Highlights

  • Benchmarked Random Forest regression directly against the classical Gilbert empirical correlation on the same held-out data
  • Built the dataset from 1,000+ daily observations across 7 wells and 4 flow stations, standardizing units and column names across sources
  • Cleaned for real field conditions — sensor errors, shut-in periods, workovers and maintenance windows were flagged and removed rather than silently averaged in
  • Tuned with GridSearchCV and analysed feature importance to identify which production variables actually drive the prediction
  • Ran residual error analysis across six sequential notebooks, from data cleaning through to multi-well comparison

Technical Challenges

Convincing domain experts that ML can outperform an established petroleum engineering formula required rigorous benchmarking rather than a headline accuracy number. The residual analysis was what did it: Gilbert's correlation showed systematic error in specific operating regimes that the Random Forest captured. The harder data problem was upstream — a mature field spends a lot of time not producing, so shut-ins, workovers and sensor faults had to be excluded deliberately, with every preprocessing decision documented. Left in, they teach the model that zero flow is normal.