A CNN-RNN based machine learning model for carbon storage management

Abstract:

The current state of reservoir management strategies depends on disparate sets of reservoir simulation models targeted at active exploration sites. Lengthy simulation time and heavy reliance on the interpretation of subject matter expertise limit their wide applications. The successful development of a machine learning (ML) virtual learning environment (VLE) helps stakeholders to learn reservoir behaviors and can accelerate the process of exploring strategies to optimize reservoir development, management, and monitoring. As part of an effort to develop such a VLE, a CNN-RNN based ML model, particularly a ConvLSTM model, has been designed to evaluate its feasibility and performance for rapid forecasting strategies using a representative set of reservoir types. The model estimates the pressure, CO₂ saturation, and water production based on injection rates, permeabilities, and porosities. In this paper, we discuss the ConvLSTM model structures suitable for subsurface predictions under different spatial and temporal domains. For the spatial domain, permeabilities, porosities, pressure, and CO₂ saturation vary in 2D and multi-layer-2D while CO₂ injection and water production are performed at specific well locations. Emphasis is put on how to refine the model based on a multi-scale concept to make accurate predictions involving singularities in the spatial domain (e.g., water production at discrete sites). For the temporal domain, we explore the effects of time-step prediction and entire life-span profile prediction. The two approaches might produce similar results, but they could potentially reflect solutions for different problems. The accuracy and performance of the model are evaluated against two toy-model data sets. The data sets simulated effects of carbon dioxide (CO₂) sequestration with an Equation-of-State (EoS) reservoir simulator, GEM, developed by the Computer Modeling Group. The results are presented together with the lessons learned in applying the CNN-RNN ML model to subsurface applications.