Quantitative evaluation of fracture porosity from dual laterlog based on deep learning method

Abstract

Fracture porosity is one of the key parameters for characterizing fractured reservoirs. However, fracture porosity calculation is difficult with conventional logging data due to severe anisotropy of the reservoirs. To deal with the problem, the equivalent macroscopic anisotropic formation model based on dual laterolog (DLL) data is adopted to cyclically assign such parameters as bedrock resistivity (RB), fluid resistivity in fractures (RFL), fracture dip angle (FDA) and fracture thickness as well as fracture spacing, and to produce massive data for formation modeling. A large number of training data obtained through three dimensional finite element forward modeling and the functional relationship between DLL responses and fracture parameters that are trained and summarized by deep neural network, are combined to establish a new fast forward model for calculating DLL responses in fractured formations. A new fracture porosity inversion model for fractured reservoirs based on gradient optimization inversion algorithm combined with multi-initial inversion strategy is then proposed. While running the model, formation is divided into eight intervals according to bedrock resistivity and fracture dip angle from 0° to 90° is divided every 0.5° to improve the operation speed and efficiency. The results of numerical verification show that when bedrock resistivity is greater than 1000 Ω m, the mean absolute error (MAE) of fracture porosity inversion is 0.001658% for horizontal fractures, 0.00413% for intermediate fractures and 0.0027% for quasi-vertical fractures. When bedrock resistivity is between 100 Ω m and 1000 Ω m, MAE of fracture porosity inversion is 0.003% for horizontal fractures, 0.0034% for intermediate fractures and 0.00348% for quasi-vertical fractures. Fracture parameters determined by the fracture porosity inversion model with actual data are in good agreement with the results of micro resistivity imaging logging.