Prediction the Equivalent Fracture Network Permeability Using Multivariable Regression Analysis and Artificial Neural Networks

Abstract

Two major steps in the preparation of static reservoir models in naturally fractured reservoirs are fracture network mapping and “upscaling” (converting) the discrete fracture network and its properties, especially permeability, into the parameters which are essential to run reservoir flow simulators. This study presents a new, practical approach to estimate the equivalent fracture network permeability (EFNP) using two different methods (i.e. multivariable regression analysis (MRA) and artificial neural networks (ANN)). Different statistical and fractal characteristics of twenty natural fracture patterns collected from the outcrops of geothermal reservoirs were measured. They were then correlated to the EFNP using MRA and several empirical equations with different values of variables proposed. Next, synthetic fracture networks were generated based on different combinations of fracture length, density and orientation, and their different statistical and fractal characteristics were measured. The EFNPs of these synthetic fracture networks were predicted, using the derived equations to validate the equations. The actual EFNPs in all of these exercises were obtained using a commercial discrete fracture network modeling simulator.
As a final effort, the capability of ANN to improve the correlations obtained through the MRA was exploited. It was shown that a back propagation (BPP) network is capable of being used as a predictive tool to predict EFNP properly.