Toward Deployable Machine Learning Virtual Flow Metering: A Reproducible Volve-Field Benchmark on Recalibration and Validation Design

Virtual flow metering (VFM) infers the production rate of a flowing well from routinely measured variables, offering a low-cost complement to physical multiphase meters and periodic well testing. Machine-learning (ML) VFMs are widely reported with high accuracy, yet most studies use proprietary data and random train-test splitting, which limits reproducibility and can overstate deployable accuracy. Using the openly licensed Volve field dataset, a reproducible VFM benchmark is built and the evaluation protocol is treated as a primary experimental variable. After cleaning and a minimum-history inclusion criterion, four producing wellbores (۷,۴۳۷ daily records, ۲۰۰۸-۰۲-۱۲ to ۲۰۱۶-۰۹-۱۷) are analysed. Daily oil rate is predicted with a re-tuned Gilbert-type correlation, linear regression, k-nearest neighbours, random forest, and histogram-based gradient boosting, under three protocols: random splitting, static (frozen) temporal splitting, and a deployment-realistic walk-forward scheme in which the model is periodically recalibrated as new data arrive. The wellhead pressure and the choke differential pressure are shown to be almost perfectly collinear (r = ۰.۹۹), and the redundant variable is removed. Random splitting gives an optimistic gradient-boosting R² of ۰.۹۸, but on identical test points a frozen model achieves only R² = ۰.۸۰ (MAPE ۶۶%), whereas the recalibrated walk-forward model—trained in logarithmic space on a pooled multi-well set with physically motivated trend features and a periodic well-test anchor—reaches R² = ۰.۹۳ (RMSE ۳۳۳ Sm³/d, MAPE ۱۸%). Per-well MAPE is below ۲۰% for all four wells. Recalibration is the single most important factor. It is concluded that a deployable ML VFM is best framed as a continuously recalibrated model for producing wells, evaluated by walk-forward testing and reported with per-well metrics

کلیدواژه ها:

Virtual flow metering ، Machine learning ، Production rate prediction ، Volve field ، Walk-forward recalibration ، Model validation