Abstract
The paper describes the approach to modelling a discrete fracture network (DFN) for a gas field in the Republic of Uzbekistan. The Jurassic productive reservoir is confined to hanging wall block in a reverse faulting environment and composed of naturally fractured low porosity carbonate rock. The developed conception of the reservoir fracturing guides integrating all fracture related data in the DFN static model, which is the essential tool for consistent design of the field development. The approach has been developed to correlate the resistivity micro imagers’ fracture density with data from "conventional" logging. The neural network is designed with appropriate parameters (number of layers, neurons, learning rate, etc.) and then trained on datasets of micro imaged wells. The population of the earth model with fracture characteristics is controlled by the complex trend which combines seismic AVAZ-attributes and structural aspects. An ensemble of the DFN realizations has been generated varying distributions of fracture geometry and aperture. The matching of the dual porosity simulation model provided a clue for selecting the appropriate realizations. Fracturing of the reservoir is caused by tectonically induced deformation of the formation. The majority of fractures belong to the zones of major folding axes and curvature peaks of the structure. The fractures’ strikes generally correspond to the directions of those axes. The trained neural network establishes the correlation of "fracture" logging data and "conventional" logging data with R2 factor of >0.90. After ensuring the absence of evident overfitting this value is considered as high enough for justifying the approach implementation. The fractures dip azimuths are defined on the circular "never-ending" range of angles, so the correct treatment of this kind data requires application of vector summation rules at every stage of modelling from well data scaling-up to the DFN generation. To avoid irrational averaging of dispersed dip azimuths within a grid cell, the pile of interpreted conductive fractures for each micro imaged well should be divided into appropriate number of independently modelling sets. Since fracture density data is of heteroscedastic nature, it cannot be distributed in the earth model using only variogram based geostatistical methods. Therefore, it is crucial to establish a trending property for a strict control of a fracture density spatial distribution within the model grid. The neural network approach enables the generalization of a limited number of micro imagers’ fracture densities to a sufficiently wider range of wells characterized solely by "conventional" logging datasets. Since fracture density spatial distribution is usually subject to abrupt lateral changes, any insights on additional independently obtained well fracture density data can significantly enhance the quality of the DFN model.
