Rapid history matching of petroleum production from well logs and 4D seismic via Machine Learning techniques in the Norne Field, offshore Norway

Abstract

Predicting oil and gas reservoir behaviour requires multiple property modelling using various formulae, relationships and empirical techniques, which is time-consuming and often ineffective in precisely capturing non-linear dependencies. Artificial Intelligence (AI) or Machine Learning (ML) techniques can build time-series models for modelling dynamic reservoir properties such as water, oil and gas saturation, and pressure, thus capturing changes caused by hydrocarbon production. Here, 4D (time lapse) seismic surveys have been used to model the changes in water saturation using AI techniques such as multi-linear regression, multi-variable kriging and random forest. Statistical testing of the resulting 3D reservoir models using R-Squared, RMSE (root-mean-square error) and MAPE (mean absolute percentage error) indicated the random forest technique gave the best results and stratification had a negligible effect. Increasing the training size set from 10% to 80% improved the statistics as expected though the rate of improvement falls rapidly above a training dataset size of 40%. This indicates that 3D models with good accuracy could be obtained even with limited data. Similar techniques can be run to build 3D time-series pressure models and the results can be used for improved history matching, forward estimation of production data and estimation of reserves.