Application of Deep Learning for Characterization of CO2 Leakage Based on Above Zone Monitoring Interval (AZMI) Pressure Data

Abstract

CCS is an important component of worldwide strategy to tackle climate change. The CO2 will be captured and stored in subsurface reservoirs for a near future to help reducing the greenhouse gas emissions to the atmosphere. To ensure that the CO2 is retained within the reservoir, proper monitoring of CCS sites is necessary. Such monitoring is important for regulators, public perception, the environment and safety. Here we present a framework that uses deep learning methodology, based on reservoir simulations and information from Above Zone Monitoring Interval (AZMI) wells, to detect, localize and quantify potential CCS-leakages trough a Multi-Task-Learning (MTL) framework. To simulate subsurface CO2 movement and its effects on the pressure in the AZMI, we use a high fidelity reservoir model. In particular, in this study we use a heterogeneous 2D spatial reservoir model to simulate different leak fluxes and locations. The work presented here is a short summary of the work presented in the paper A Variational Auto-Encoder for Reservoir Monitoring available on arxiv.

Keywords: AZMI; Pressure Reconstruction; Classification; Bayesian Variational Methods, Semi Conditional Variational Auto-Encoders; Multi tasks Learning; CCS Monitoring