Data-Driven Models to Predict Hydrocarbon Production From Unconventional Reservoirs by Thermal Recovery

In the numerical simulations of thermal recovery for unconventional resources, reservoir models involve complex multicomponent-multiphase flow in non-isothermal conditions, where spatial heterogeneity necessitates the huge number of discretized elements. Proxy modeling approaches have been applied to efficiently approximate solutions of reservoir simulations in such complex problems. In this study, we apply machine learning technologies to the thermal recovery of unconventional resources, for the efficient computation and prediction of hydrocarbon production. We develop data-driven models applying artificial neural network (ANN) to predict hydrocarbon productions under heterogeneous and unknown properties of unconventional reservoirs. We study two different thermal recovery methods—expanding solvent steam-assisted gravity drainage for bitumen and in-situ upgrading of oil shale. We obtain training datasets by running high-fidelity simulation models for these two problems. As training datasets of ANN models, diverse input and output data of phase and component productions are generated, by considering heterogeneity and uncertainty. In the bitumen reservoirs, diverse permeability anisotropies are considered as unknown properties. Similarly, in the oil shale reservoirs, diverse kerogen decomposition kinetics are considered. The performance of data-driven models is evaluated with respect to the position of the test dataset. When the test data is inside of the boundary of training datasets, the developed data-driven models based on ANN reliably predict the cumulative productions at the end of the recovery processes. However, when the test data is at the boundary of training datasets, physical insight plays a significant role to provide a reliable performance of data-driven models.