Multiple fractures or an extensive fracture network iscritical for commercially viable production from lowpermeability formations, such as shales. Mechanicalanisotropy is inherent in shales because of its platy nature.This inherent anisotropy makes fracture prediction in shalesmore complex, and traditional methods to predict fracturegeometry assuming isotropy frequently prove to beinadequate. Current analytical methods boldly assume aconstant fracture height and constant mechanical propertiesfor the entire height. Common 3D fracture modelingsoftware are based on isotropic rock models, and modelsthat take anisotropy into account are computationallyexpensive and time consuming, especially when numeroussimulations must be performed by varying the inputparameters for parametric study.This paper proposes a workflow to improve the predictionof fracture geometry in anisotropic formations. Theworkflow involves generating a neural network by using alimited number of 3D fracture modeling cases. After theneural network is obtained from a pilot or offset well, it canbe easily embedded into software for optimizing fracturedesign, identifying geologic sweet spots, and predictingfracture propagation and correlating the results to otherhorizontal or vertical wells in the same geological area.This process can be divided into three steps. First, theanisotropic models are used to predict horizontal andvertical Young’s modulus (E_(horz) and E_(vert)), Poisson’s ratio(v_(horz) and v_(vert)), and anisotropic minimum horizontal stress(σ_(hmin_ani)) from sonic and density log measurements. Second,the elastic moduli properties and σ_(hmin_ani) are entered into a3D fracture modeling simulator to run different cases byvarying the completion input parameters. The outputs of thefracture simulator (i.e., the fracture length, height, width,and effective length) serve as a training database to theneural network. In the final step, a neural network isgenerated based on the training database. After thereservoir-specific neural network is developed, fracturegeometry can be predicted or optimized for numerouscombinations of completion input parameters in a timelyand cost effective manner. Because the commonly availablecommercial fracture modeling software assumes isotropy, anew method is presented in this paper to representmechanical property anisotropy using equivalent Young’smodulus (E_(eq)) and Poisson’s ratio (v_(eq)). E_(eq) and v_(eq) arederived from E_(horz), E_(vert),v_(horz), and v_(vert) and the isotropic (Sneddon and Berry 1958) and anisotropic (Chertov 2012)width functions.This workflow is demonstrated by generating a neuralnetwork for two reservoirs using anisotropic elastic modulias predicted by the dipole sonic log. The fracture geometrypredicted by the neural network is compared with theconventional method, assuming the isotropic shale rock. Theresults show that by assuming an isotropic model thefracture width is overestimated, and the fracturecontainment and propped length are underestimated. Theanisotropic neural network model is further run in a largeparametric study to demonstrate how the effective lengthvaries with perforation position, injection volume, andinjection rate. The results helped to optimize perforationdepth, injection rate, and pumped volume.
展开▼
