Development of a Data-Driven Operational Design Tool for CO2 Sequestration in Shale Gas Reservoirs

Abstract

Long-term sequestration of industrial CO2 in geological formations, such as shale, has been gaining interest to reduce the global greenhouse effects and its side effects on climate change. In this study, a data-driven operational design tool for CO2 sequestration in shale-gas reservoirs is developed and tested. The model described is based on artificial neural networks trained with a large number of numerical-simulation scenarios of natural gas production and CO2 injection. Numerical simulations were performed using PSU-SHALECOMP, a compositional dual-porosity, dual-permeability, multi-phase reservoir simulator developed at Penn State University. The simulator also incorporates the effects of water presence in the micropore structure and those of matrix shrinkage and swelling. The tool was designed in an inverse- looking fashion to estimate the necessary wellbore and hydraulic-fracture design characteristics in the form of stimulated-reservoir-volume (SRV), once known initial conditions, reservoir rock and fluid characteristics, and desired long-term natural gas production and CO2 injection profiles are input. During the gas-production period, the well is operated with a flowing bottom-hole pressure constraint. CO2 sequestration is performed with a constant injection rate, until a specified fracturing-pressure limit is reached. The profiles are input in the form of 12 equally-divided time steps, which are also specified explicitly as input parameters. After running a large number of different scenarios, an artificial neural network was trained and validated using the dataset obtained from simulation runs. To determine the optimum neural- network architecture, almost 100,000 different neural-network designs were tested with parallel-processing in UNIX clusters. Blind-testing of cases resulted in an average prediction error of 9.8% for output parameters, which would help to accurately design the wellbore and SRV characteristics needed for the planned sequestration process. Re-validating the testing cases using the numerical model resulted in an average error of 3.2% for cumulative gas production and an average error of 8.6% for cumulative CO2 injection. A graphical-user- interface application was also developed that enables using the model in a practical and efficient manner. Such kind of a tool would help operators to design CO2 sequestration projects in shale gas reservoirs in an efficient manner.

AAPG Datapages/Search and Discovery Article #90335 © 2018 AAPG 47th Annual AAPG-SPE Eastern Section Joint Meeting, Pittsburgh, Pennsylvania, October 7-11, 2018