Virtual Flow Metering: Forecasting Oil Production with Machine Learning

Natalia Kovalchuk, Muhammad Arsalan24 May 2025

Problem

An oil producer needs the oil, gas and water flow rate (debit) of every well in real time. Under-delivery to a refinery brings penalties and a plant stop, and over-delivery "just in case" is lost profit. A rising water share can make a well unprofitable. The usual tools fall short: a separator is accurate but slow, a multiphase flowmeter is complex and expensive, and a mobile flowmeter visits a well about once a day and needs the well shut down.

Example from the article: every well already has pressure, temperature and choke-size sensors, and the task is to learn Q=Q(P1,T1,P2,T2,P3,T3,Choke\%,…) for each phase.

Solution in brief

Treat it as tabular regression from cheap sensors to each flow rate, not as time series. Clean new data automatically with physical checks, Z-scores and autoencoders. Build features from physical laws, select them in two steps, and use a separate pipeline per phase. Water gets a two-stage pipeline (low/high classifier, then a regression per class) because high water rates are rare.

Questions this case answers
  • How to predict a unit's output rates from cheap sensor readings when the target has very few high values and experts can propose more features than there are data points
  • How to build an automatic cleaning step and features grounded in physical laws for a model that engineers will run on new data themselves
  • How to estimate oil, gas and water flow rates of a well from pressure, temperature and choke telemetry without a multiphase flowmeter
  • How to build an automatic data-cleaning step for engineers who will run the model on new data themselves

Tools and methods

  • Physical validity checks, Z-scores, autoencoder reconstruction error — automatic cleaning of new telemetry
  • Physics-based feature engineering — Bernoulli, Darcy and Gilbert equations
  • Concrete Autoencoder, feature importance, wrapper methods — stepwise feature selection
  • Classifier plus per-class regressions with covariate shift reduction — pipeline for rare high rates
  • Gradient boosting, MLP — models; MLP is slightly more accurate and retrainable
Continue reading — it's free

Google or a one-time email code. No password.