Time-Lapse Seismic Matching for CO2 Plume Detection via Correlation-Based Recurrent Attention Network

Time-lapse seismic data analysis is an effective technique for monitoring reservoir changes and plays an important role in CO2 plume detection. Theoretically, during CO2 geological storage, time-lapse seismic data corresponding to nonreservoirs should be consistent. However, due to changes in near-surface velocity, differences in the position of sources and receivers, and different acquisition parameters and instruments, the consistency of time-lapse data is poor, which seriously affects the application of time-lapse data. To solve this problem, a correlation-based recurrent attention network (CRAN) is proposed for consistency matching of time-lapse data. Cross correlation is a direct measure of the correlation between seismic traces. The improved loss function of the network based on cross correlation is beneficial for improving the matching accuracy. Recurrent neural networks (RNNs) can learn sequential relations and suitable for predicting time series. Therefore, RNNs are combined with convolutional neural networks to improve the network time sensitivity. Treating a large number of extracted features equally is not conducive to effective information recognition; thus, a channel attention mechanism is introduced to assign different weights to the features, thus improving the contribution of useful information. The test results of the synthetic and field data show that the proposed CRAN has an excellent enhancing effect on the repeatability of time-lapse data, and the matched data clearly reveal a plume of CO2.

Automated summary

Time-lapse seismic data analysis is important for monitoring changes in reservoirs and detecting CO2 plumes. In order to address the inconsistency issue in time-lapse data caused by various factors, a correlation-based recurrent attention network (CRAN) is proposed. The network utilizes cross correlation to calculate the correlation between seismic traces and introduces a channel attention mechanism to assign weights to different features. The results of testing on synthetic and field data demonstrate that CRAN significantly enhances the repeatability of time-lapse data and accurately reveals CO2 plumes.