A Hybrid Artificial Intelligence Framework with Residual Correction for Predicting Oil Well ProductivityA Hybrid Artificial Intelligence Framework with Residual Correction for Predicting Oil Well Productivity

Abstract

The unpredictability of overall productivity for an oil well has remained troublesome because of the non-linear behaviour of reservoirs, as well as issues associated with multi-phase flow and elevated levels of uncertainty. Conventional means of predicting productivity (e.g., Darcy’s law, inflow performance relationship curves, and numerical simulations) typically require large amounts of historical production data, are time-consuming or are not necessarily accurate for changing field conditions. A Unique Artificial Intelligence (AI) Based Predictive Framework was created to provide real time productivity state estimates and project future production rates for an oil well. The framework was built using a set of field data from over 500 wells, which consisted of reservoir characteristics, completion characteristics, PVT properties and production history. Data needed to be pre-processed through normalization and outlier removal techniques. Various modelling techniques including Artificial Neural Networks (ANN), Random Forest (RF), XGBoost (XGB) and the proposed hybrid modelling technique (XGB with a residual and stacking approach using Bayesian optimization) were compared. Of these modelling techniques, the proposed hybrid technique achieved a greater than 0.94 R² and demonstrated a 15-20% reduction in Root Mean Square Error (RMSE) when compared to conventional A.I. modelling. The analysis of the feature importance revealed that both permeability and skin factor are the most significant features for estimating oil well productivity. The result of this study enables productivity projections to be made in real-time, provide information to assist in well intervention activities through identifying stimulation (intervention) targets, and generate geological insights for use in reservoir engineering applications. The unique predictive framework is transferable to other similar oil fields and provides a flexible and interpretable tool for reservoir engineers.