TY - THES

T1 - Research on Deep Learning-based Microgravity Surveillance Method for CO2 Storage in Deep Saline Aquifers

AU - Ao, Ran

N1 - no embargo

PY - 2023

Y1 - 2023

N2 - In China's pursuit of Carbon Peak and Carbon Neutrality, geological storage of carbon dioxide in deep saline aquifers has become a means of underpinning China's energy sector to achieve its dual carbon targets. With the gradual expansion of storage projects worldwide, there is an urgent need for an efficient and convenient monitoring method that can provide long-term monitoring of underground CO2 storage. Microgravity monitoring observation is efficient, and the measurement cost is low. It can obtain time-lapse gravity anomalies by high-precision gravimetry, estimate CO2 subsurface density distribution, and delineate the CO2 injection and plume boundary. Traditional gravity anomaly inversion methods have low depth resolution, scattered inversion results, time-consuming computation of large data volumes, and high storage requirements when dealing with geological situations where the horizontal scale of CO2 is much larger than the vertical scale. In contrast, deep learning inversion methods observe from a probabilistic perspective and control the minimization of network output and model labels through loss functions to enable supervised learning-based neural network training. However, the existing model label datasets, including MNIST and GravInv, significantly differ from the measured microgravity anomaly response of CO2 geological storage. Some individual studies manually create accurate models that match the gravity anomaly response but cannot meet the data volume requirements of the neural network in terms of the number of model labels. In this paper, a dynamic density geological model of CO2 sequestration that meets both the microgravity anomaly response and the neural network learning quantity requirements is constructed based on the geological conditions of CO2 storage, and the neural network built and trained for this geological model label is applied to the inversion of the measured data. The geological model labels are constructed and trained for the inversion of the measured data. The input field data are used to predict the subsurface CO2 density distribution. The deep learning inversion results show that high lateral and vertical resolution can be achieved in areas of density change, and the top-bottom interface and horizontal boundary of the CO2 reservoir can be identified. During the neural network¿s learning process, the dynamic density structure and horizontal spreading trends of the CO2 geological model are identified, reducing the incorporation of other prior information in the inversion process. The completed neural network can fast real-time inversion, significantly reducing computation time compared to conventional gravity inversion methods. In the inversion of time-lapse microgravity anomaly data from the measured Sleipner field, the multilayer structure and enrichment areas of CO2 were predicted, and agreement was achieved with other studies, validating the practical effectiveness of the inversion network. A promising application of deep learning-based monitoring of CO2 geological storage microgravity anomalies is displayed.

AB - In China's pursuit of Carbon Peak and Carbon Neutrality, geological storage of carbon dioxide in deep saline aquifers has become a means of underpinning China's energy sector to achieve its dual carbon targets. With the gradual expansion of storage projects worldwide, there is an urgent need for an efficient and convenient monitoring method that can provide long-term monitoring of underground CO2 storage. Microgravity monitoring observation is efficient, and the measurement cost is low. It can obtain time-lapse gravity anomalies by high-precision gravimetry, estimate CO2 subsurface density distribution, and delineate the CO2 injection and plume boundary. Traditional gravity anomaly inversion methods have low depth resolution, scattered inversion results, time-consuming computation of large data volumes, and high storage requirements when dealing with geological situations where the horizontal scale of CO2 is much larger than the vertical scale. In contrast, deep learning inversion methods observe from a probabilistic perspective and control the minimization of network output and model labels through loss functions to enable supervised learning-based neural network training. However, the existing model label datasets, including MNIST and GravInv, significantly differ from the measured microgravity anomaly response of CO2 geological storage. Some individual studies manually create accurate models that match the gravity anomaly response but cannot meet the data volume requirements of the neural network in terms of the number of model labels. In this paper, a dynamic density geological model of CO2 sequestration that meets both the microgravity anomaly response and the neural network learning quantity requirements is constructed based on the geological conditions of CO2 storage, and the neural network built and trained for this geological model label is applied to the inversion of the measured data. The geological model labels are constructed and trained for the inversion of the measured data. The input field data are used to predict the subsurface CO2 density distribution. The deep learning inversion results show that high lateral and vertical resolution can be achieved in areas of density change, and the top-bottom interface and horizontal boundary of the CO2 reservoir can be identified. During the neural network¿s learning process, the dynamic density structure and horizontal spreading trends of the CO2 geological model are identified, reducing the incorporation of other prior information in the inversion process. The completed neural network can fast real-time inversion, significantly reducing computation time compared to conventional gravity inversion methods. In the inversion of time-lapse microgravity anomaly data from the measured Sleipner field, the multilayer structure and enrichment areas of CO2 were predicted, and agreement was achieved with other studies, validating the practical effectiveness of the inversion network. A promising application of deep learning-based monitoring of CO2 geological storage microgravity anomalies is displayed.

KW - Kohlendioxid

KW - geologische Speicherung

KW - Zeitraffer-Mikrogravitation

KW - Deep Learning

KW - neuronale Netze

KW - carbon dioxide

KW - geological storage

KW - time-lapse microgravity

KW - deep learning

KW - neural networks

M3 - Master's Thesis

ER -