In-situ stresses prediction by using a CNN-BiLSTM-Attention hybrid neural network

ABSTRACT: In-situ stresses prediction plays an important role in petroleum exploitation, such as wellbore stability analysis and hydraulic fracturing design. However, due to the complexity, uncertainty and heterogeneity of deep formation, there is a strong nonlinear relationship among different logging data, which makes it difficult to express the spatial correlation of the logging data using analytical model and traditional neural network model. Therefore, based on the logging data of Da’anzhai shale formation in Sichuan Basin, a hybrid neural network was proposed to predict the in-situ stresses. This hybrid neural network involved the Convolutional Neural Network (CNN), the Bidirectional Long Short-Term Memory neural network (BiLSTM), and the Attention mechanism by integrating the advantages of CNN, BiLSTM, and Attention mechanism, and it is therefore called as CNN-BiLSTM-Attention hybrid neural network. The logging data of Da'anzhai shale formation in N-2 well was utilized to verify this CNN-BiLSTM-Attention hybrid neural network. The results indicated that the CNN-BiLSTM-Attention hybrid neural network can accurately predict in-situ stresses. The maximum relative error of vertical stress, maximum horizontal principal stress, and minimum horizontal principal stress are 0.94%, 1.54%, and 1.78%, respectively. Therefore, the CNN-BiLSTM-Attention hybrid neural network is recommended for predicting in-situ stresses.