Intelligent prediction of fracture parameters in ultra-deep carbonate rocks based on knowledge and data dual drive

Traditional fracture characterization techniques based on core and imaging logs are too expensive to cover and extrapolate to all wells. In this study, fracture characterization is achieved by comprehensively considering logging responses and geomechanical parameters, combined with the extreme value and variance analysis stratification method, Pearson correlation analysis, principal component analysis (PCA), and long short term memory (LSTM) neural network. The results demonstrate that fracture-developed intervals can be accurately divided by the extreme value and variance analysis stratification method and the logging response. The introduction of geomechanical parameters, Pearson correlation analysis, and PCA can achieve data dimensionality reduction while retaining the information contained in the original indicators, which directly improve the network performance. LSTM network can fully mine the information within and between data, which is highly suitable for fracture characterization. Taking the predicted fracture density as an example, the RMSE is 1.55. Compared with Backpropagation Neural Network (BP), Support Vector Regression (SVR) and Convolutional Neural Network (CNN), the RMSE of this method decreases by 0.08, 0.09 and 0.01, respectively. The findings of this study can help for rationalizing the exploration and development program of carbonate reservoirs and enhance the production of hydrocarbons in naturally fractured reservoirs.