| Popis: |
An important part of carbon capture and storage (CCS) Project is geophysical monitoring of carbon dioxide (CO2) injection in deep saline aquifers. The durability of this type of sequestration depends entirely on the long-term geological integrity of the seal. There is a strong correlation between the change in water saturation and the change in CO2 saturation in a saline reservoir. Electromagnetic monitoring technology based on electromagnetic inversion has been widely used as a powerful supplement to expensive seismic methods in the CO2 sequestration monitoring field. Dissolved salt reacts with the CO2 to precipitate out as carbonates thereby changing the complex electrical resistivity. So, there is a direct correspondence between the change in saturation and the measured electric field on the ground surface, which makes electromagnetic (EM) methods very suitable for monitoring CO2 sequestration. Based on electromagnetic inversion, the monitoring method attempts to estimate the distribution of CO2 in the deep reservoir from observations on or above the ground to locate the plume position. In order to improve the resolution and detection depth of electromagnetic detection, we have developed a new method of joint inversion of electromagnetic induction polarization with Nano-enhancement technology and casing excitation technology. Inverse problems are usually posed as least-squares optimization problems in high-dimensional parameter spaces. The existing approache based on deterministic gradient-based methods is limited by the nonlinearity and no-uniqueness of the inverse problem. Probabilistic inversion method, despite its great potential in uncertainty quantification, still remains a formidable computational task. In this work, we explored the potential of deep learning methods for borehole-to-surface electromagnetic (BSEM) inversion. This method has a natural advantage that does not require calculation of gradients and provides immediate results. Deep convolutional neural network based on U-net architecture is trained on large synthetic datasets obtained by our full 3-D BSEM Code. This method has been verified on the practical model of the typical onshore borehole-to-surface controlled source CO2 monitoring scenario. Pre-trained networks can reliably estimate the location and lateral dimensions of the CO2 plume, as well as their complex resistivity properties. We also compared several fully convolutional network architectures in terms of generalization, accuracy, timeliness and training cost. The feasibility of the deep learning inversion has been further confirmed from examples with different survey geometry and noise, which opens the possibility to estimate the complex resistivity distribution of the Deep CO2 sequestration in real time. |
