Modeling Naturally and Hydraulically Fractured Reservoirs with Artificial Intelligence and Assisted History Matching Methods Using Physics-Based Simulators

Authors:

Abstract
Unconventional reservoirs typically exhibit large uncertainty in rock and fracture properties coupled with significant heterogeneity making manual history matching a challenging endeavor. Computer-assisted methods of history matching gained preference because they honor the non-uniqueness of the possible geological and fracture realizations by allowing efficient simulation sensitivities over many possible outcomes. This study integrates EDFM's fracture-modeling capabilities into an artificial intelligence-based history matching and optimization tool, to achieve more confidence and capability in simulations for production forecasting and reservoir-behavior studies. This approach involves four main steps: (1) initializing EDFM-fracture parameters for a realization, (2) running a pre-simulation command that calls the EDFM-preprocessing engine and generates an updated model, (3) running the model on an appropriate simulator, and (4) calculating the objective function. We apply this method to perform history matching for a horizontal tight-oil well in the middle Bakken. We further repeat history matching for the same model after including a set of 1,000 randomly distributed natural fractures. In both studies, four embedded-fracture parameters are set to vary within predefined continuous ranges. The objective is to minimize history matching error for field-recorded, bottom-hole pressure and gas production data with oil production history set as a constraint. This new method, with a more realistic fracture geometry, resulted in a better match to the field-recorded data compared to a previous study that was based on a local gird refinement model. After introducing natural fractures to the model, history-matching results displayed a trend of decreasing rock-matrix permeability and fracture conductivity compared to the original scenario without natural fractures. In this case, history matching error response was most sensitive to rock-matrix permeability followed by fracture half-length. We then made 30-year production forecast using the most optimal solutions. Accepted solutions were screened based on a cut-off value of 5% normalized history-matching error. We finally demonstrate that, even though the presence of natural fractures significantly altered the model, it did not have a major influence on long-term production.

This publication has 21 references indexed in Scilit: