Improved Correlations and Predictive Models for Nigerian Crude Oil Pvt Properties Using Advanced Regression and Intelligent Techniques

Abstract Calculating reserves in an oil reservoir and also the determination of its performance and economics needs a sound knowledge of the fluids’ physical properties. Bubblepoint pressure, Oil FVF at bubblepoint, solution GOR and compressibility are of primary importance in material balance calculation. Ideally, the laboratory measurements of PVT properties are the primary sources of PVT data gotten from laboratory studies of the samples collected from the wellbore or from the surface. However, such experimental data are not always available mainly due to unreliability of the data from samples or that samples have not been taken as a result of cost saving as the operation can be quite expensive. It therefore became pertinent to be able to predict these properties in the absence of the experimental data. Several correlations have been developed to enhance this prediction but, unfortunately, these correlations are highly region dependent and rarely works well in different regions since they were developed with data from particular regions. The objective of this study is to develop better performance correlations and predictive models for the Nigerian crude oil PVT properties using linear and non-linear multivariate regression techniques and supervised machine learning techniques. This study focuses on predicting bubblepoint pressure, Pb, oil formation volume factor at bubblepoint, Bob, and solution gas-oil ratio, Rs, as functions of the reservoir temperature and the oil and gas gravities. In this study, evaluation and tuning/recalculation of coefficients (using linear and non-linear regression analysis) of some of the best PVT correlations to estimate the desired PVT parameters and employing the KNN and the Random Forest algorithms to develop better PVT models for the Niger Delta region, were performed. After adequate pre-processing, the gathered dataset was divided into training and test datasets by random sampling. Average absolute Relative Error, Root Mean Square Error and the Correlation coefficient, R are the main loss functions employed to evaluate and compare the developed models with the conventional correlations. The new models showed much better performance as regards the Niger Delta crude PVT properties. With respect to the best performing conventional correlations for the Nigerian crude, the new models decreased the RMSEs by 72%, 61.1% and 83% and increased the correlation coefficients by 35%, 7.2% and 26.9% for the bubblepoint, oil formation volume factor and solution gas-oil ratio respectively.