Permeability prediction of heterogeneous carbonate gas condensate reservoirs applying group method of data handling

Abstract

Carbonate petroleum reservoirs typically have lower permeabilities and recovery factors than sandstone reservoirs, so the natural fractures they often incorporate have positive impacts on resource recovery and fluid production rates. Quantifying effective permeability, incorporating contributions from pores and fractures, is therefore essential in the reservoir characterization and flow-regime modelling of carbonate reservoirs. This research applies a robust machine-learning forecasting model to predict permeability (K) for heterogeneous carbonate gas condensate reservoirs. A 212-point dataset from six gas-condensate carbonate reservoirs (Russia and Iran) is compiled. The input variables considered are porosity (Φ, %), specific surface area (Sp, 1/cm) and irreducible water saturation (Swir, %). These variables are assessed using four machine learning models: group method of data handling (GMDH), polynomial regression (PR), support vector machine (SVR), and decision tree (DT) to predict permeability. The GMDH algorithm, a polynomial neural network with a customized architecture is developed, such that it displays increased prediction accuracy and improved learning capabilities. All four models developed in this study substantially improve upon K predictions derived from established empirical correlations. The GMDH model also outperforms the other models in respect of K prediction accuracy using Φ, Swir, and Sp as input variables. It achieves permeability prediction accuracy for the multi-field dataset evaluated with a root mean squared error (RMSE) and coefficient of determination (R2) for the training and testing of the best model (GMDH) of RMSE = 9.2 mD and R2 = 0.9988; RMSE = 0.4 mD and R2 = 0.9972, respectively. The model can be readily adapted for application to other field datasets to estimate K from limited well-log and/or core data.