A machine learning-based, direct prediction framework for timely forecasting of CO2 plume migration
Abstract
Accurate and timely forecasts of CO2 plume evolution in geological storage formations are crucial for the assessment of CO2 injection risks, permanent underground sequestration, and safe and efficient management of commercial-scale deployments. Conventional forecasting often involves a two-step strategy: first incorporate monitoring data to calibrate reservoir model parameters and then run the forward model to predict CO2 plume distributions. This method impedes in-time leakage detection and real-time decision-making for CO2 storage site operators due to the large computational demand in performing commercial-scale reservoir simulations and the difficulty in constraining reservoir parameters using the limited observations. In this work, we propose a machine learning-based, direct forecasting framework to accelerate the prediction of CO2 plume migration by learning the observation-prediction relationship directly without the model inversion. We first map the high-dimensional plume extent onto a low-dimensional latent space using a convolutional autoencoder. We then use neural networks to learn the relationship between the low-dimensional prediction variables and the observation variables. Lastly, we infer the prediction values directly from the observation data. The autoencoder provides a good parameterization of the CO2 plume and facilitates the subsequent relationship learning. The direct prediction from the neural network evaluation greatly improves the computational efficiency making the real-time forecasting feasible. We apply the proposed method to predict the CO2 plume using the limited observation data in a monitoring well. Results indicate that it can accurately forecast the CO2 plume area, centroid movement distance, and plume spreading in the primary and secondary directions, which demonstrates its effectiveness in predicting the spatio-temporal evolution patterns of plume migration under diverse geological complexities. This one-step direct forecasting is computationally efficient which requires a small number of parallelizable reservoir simulations and it can provide accurate predictions with limited observations by learning the observation-prediction relationship in the reduced dimension.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H11J..07F