Application of Machine Learning Ideas to Reservoir Fluid Properties Estimation

Abstract

Abstract Machine learning refers to a range of data-driven techniques that give computers the ability to learn from exposure to data and to make predictions based on the learning. Popular applications of machine learning include hand-written digit recognition technology used by some banks to automatically process cheques, spam filtering technologies used by email applications to detect spam mails and object recognition technologies in self-driving cars, to name a few. Examples from the Oil and Gas sector, though less exotic, have also been growing steadily. For example, artificial neural networks have been used for years for the estimation of reservoir properties such as permeability and porosity; there have also been applications of the technique in the analysis of the huge amount of pressure and flow rate data from permanent downhole gauges; also, data-driven predictive analytics have been applied in mature fields with huge amounts of data. This paper discusses the results of an investigation of the performance of some machine learning techniques in the prediction of reservoir fluid properties. The techniques investigated include K Nearest Neighbors (KNN), Support Vector Regression, Kernel Ridge Regression, Random Forest, Adaptive Boosting (Adaboost) and Collaborative Filtering. PVT data from a database of 296 oil and 72 gas reservoirs from the Niger Delta were used in the study. The input data used in the training include initial reservoir pressure, saturation pressure, solution gas oil ratio (for oil samples), formation volume factor, condensate gas ratio (for gas samples), API gravity, gas gravity, saturated oil viscosity and dead oil viscosity. Trained models were developed using the techniques and used to predict saturation pressure and formation volume factor, oil viscosity and condensate gas ratio respectively for samples that were not part of the training. It was found that all six techniques gave very good results for the oil formation volume factor, comparable to and in some cases exceeding the performance of standard industry correlations such as Standing and Vasquez-Beggs. The techniques also gave good results for bubble pressure better than the standard correlations. For oil viscosity, the Random Forest and Adaptive Boosting gave very good results, of the same quality as that obtained with the popular Beggs-Robinson correlation, and did not require dead oil viscosity data. Performance of the techniques in estimating gas PVT parameters was not as good; due perhaps to the limited gas data. However, Adaptive Boosting and Support Vector Regression gave good results for dew point pressures. Overall the results indicate that these machine learning techniques offer promise for fluid properties estimation and should be given consideration where a company has acquired large amount of PVT data in a geological basin it operates.