Prediction of deep low permeability sandstone seismic reservoir based on CBAM-CNN

Abstract

The prediction of sand body is a key focus in evaluating reservoir distribution, playing a crucial role in oil and gas exploration and development. To clarify the development characteristics of the sand body in the Huangyan structural belt of the Xihu Sag and effectively identify the hidden channel, a new approach is proposed. This approach incorporates a Convolutional Block Attention Module (CBAM) into a Convolutional Neural Network (CNN). Well logging and seismic data were used to predict the distribution of sand body in the lower section of the Huagang Formation (H12). Firstly, based on known sand-stratum ratio data and well-side seismic data, the seismic attributes that are sensitive to the reservoir response are preferred through correlation analysis, which are used to construct a Polynomial Linear Regression (PLR) model to obtain the preliminary sand distribution prediction results. Secondly, the grid is divided according to this result. Then, three sample selection methods, namely proportional, fixed-range, and random, are used to construct three sample sets in conjunction with the geologic model guide. Finally, CBAM-CNN, CNN, Random Forest (RF), Support Vector Machines (SVM), and Backpropagation Neural Network (BPNN) are utilized for sand body prediction. The results indicate that when using the same model, the sample set selected by the fixed-range selection method yielded the best prediction outcome. When using the same sample set, the CBAM-CNN model outperformed the others. Among various strategies, training the CBAM-CNN model with the sample set derived from the fixed-range selection method led to the highest test set R2 of 0.913. The results of the sand body distribution indicate that fine river channels are more continuous, and reservoir boundaries are clearer, significantly outperforming other schemes. This outcome can serve as a guide for further reservoir delineation.