--> Deep Learning Used in Permeability Prediction of Channel Sand Bodies With Strong Heterogeneity
[First Hit]

AAPG ACE 2018

Datapages, Inc.Print this page

Deep Learning Used in Permeability Prediction of Channel Sand Bodies With Strong Heterogeneity

Abstract

Permeability prediction has long been one of the most important and difficult tasks of reservoir characterization. Because of strong heterogeneity, Previous HitnonlinearNext Hit relationships exist between permeability and other reservoir parameters. In order to obtain a unified permeability prediction model, many Previous HitnonlinearNext Hit machine learning methods have been used to characterize the Previous HitnonlinearNext Hit relationships. The artificial neural network has been one of these methods for many years, but its application is largely restricted because of its shortcomings, such as hard to train, overfitting, black box, and huge computational burden. However, deep learning, developed from the artificial neural networks, has overcome most of these problems with new training skill or computing method in recent years. The deep learning has been successfully applied in many fields, but its performance in permeability prediction has not been studied yet.

In this research, the core and well log data of channel sand bodies are selected from the Middle Jurassic Shaximiao Formation in the Western Sichuan Basin. The channel sand bodies are independent with strong heterogeneity, and the permeability changes significantly. It is difficult to find the Previous HitnonlinearTop prediction mode of the permeability. Based on data clearing and character analysis, we used the deep feedforward neural network (DFNN), one type of the deep learning models, in the permeability prediction problem. We tested the sensitivity of the main parameters, such as activation function, number of neurons and hidden layers, and dropout. The best network architecture and parameters are obtained for permeability prediction. According to the results of error evaluation and k-fold cross-validation, the deep learning model is more stable and accurate on blind wells, compared with other regression methods, such as linear regression (LR), multiple linear regression (MLR), and support vector regression (SVR). The deep learning model has the best fitting and generalization ability for permeability prediction with strong heterogeneity. The research indicated that the deep learning has great potential in many other complicated regression and classification problems in reservoir characterization.