The Used of an Intelligent Data Assimilation Protocol for Plume Characterization of CO2 Sequestration in Saline Aquifers

Abstract

The gas plume characterization of CO2 sequestration processes is a typical inverse problem, the solution of which could include inevitable non-unique solutions and uncertainties. This work aims at structuring a plume characterization protocol via coupling deep neural network models and ensembled Kalman Filter algorithm by analyzing injection and monitoring well bottomhole pressure data. Considering the multiple sequestration mechanisms, the output of the model includes the spatiotemporal evolution of free gas plume and mineralization profiles. In this study, one inverse model and two types of forward-looking models are developed. The inverse model predicts geological characteristics using field pressure data as input. The forward-looking models aim to simulate pressure responses and the evolution of the gas plume. Initially, input data is processed through the inverse model to estimate the spatial distribution of geological properties. The forward-looking models couple the ensemble Kalman filter to refine the predictions made by the inverse model. After the predictions are aligned with the field data, these models proceed to forecast the distribution of the gas plume in different form of presences. The proposed methodology was evaluated using an ideal case and a field case using the geological data collected from a real aquifer. The primary source of uncertainty in gas plume characterization stems from the inherent non-uniqueness of solutions to inverse problems. This methodology utilizes the heterogeneities in petrophysical properties, as predicted by the inverse model, as an intermediary variable. The incorporation of an AI-assisted data assimilation protocol substantially reduces this uncertainty by refining the outcomes from the inverse model. The findings indicate that deep neural networks models adapted from an auto encode architecture derived from the U-net are effective for image-to-image regression predictions using static heterogeneity property distributions as inputs. This approach leads to the training of an expert system capable of forecasting the varied spatiotemporal dynamics of carbon species within saline aquifers. After injection ceases and the pressure transient spreads, the gas plume evolution becomes less sensitive to changes in the injection well bottomhole pressure. It highlights the importance of strategically placing monitoring wells to effectively oversee long-term gas migration and plume characteristics. The integration of expert systems with ensemble Kalman filters successfully delineates the spatial and temporal evolution of CO2 plumes by analyzing pressure data from injection and monitoring wells. This method, in contrast to traditional CO2 plume inversion techniques, demands fewer and more affordable data inputs. As a result, it offers cost-effective and precise CO2 plume characterization considering different forms of presences.