Random Forest Ensemble Model for Reservoir Fluid Property Prediction

Abstract

Abstract Reservoir fluid PVT properties are measured in the laboratory for various use in reservoir engineering evaluation and estimation. Despite the indispensability of these PVT parameters, PVT lab data are seldomly available and if available may be unreliable. Instead, various empirical models have been developed and used in the industry. These empirical models are inherently inaccurate when used to predict PVT properties of fluid from different geological region with different depositional environment and fingerprint. Artificial Intelligence (AI) has evolved over the years and provided some algorithms with potentials to develop accurate predictive model for the prediction of bubblepoint pressure. This work tested some AI algorithms, compared performances and choose random forest regression algorithm in developing a robust predictive model for the estimation of bubblepoint pressure. Two thousand five hundred and twenty-two datasets obtained from oil reservoirs in different geographical locations were used for the feature scaling of input data, training and testing of the models. The independent variables, gas-oil ratio, temperature, oil density and gas density were confirmed to have key influence on the dependent variable Bubblepoint pressure The random forest model developed uses ensemble learning approach, combines predictions from multiple machine learning algorithms by averaging all predictions to make a more accurate prediction. The ‘forest’ generated by the random forest algorithm was trained through bootstrap aggregating. This is an ensemble meta-algorithm that improves the accuracy of machine learning algorithms. Percentage data split was 70% training and 30% testing. The reliability, accuracy and completeness of the predictive model capability were computed through performance indices such as the root mean square error (RMSE) and mean absolute error (MAE). The best network architecture was determined along with the corresponding test set RMSE, and Correlation coefficient. Statistical and graphical error analysis of the results showed that the random forest model performed better than existing models with 0.98 correlation coefficients for bubblepoint pressure. Better accuracy of reservoir properties prediction could be achieved using this random forest reservoir fluid properties prediction model.