This work proposes three new methods to capture near-fracture transient effects that occur during the simulation of unconventional reservoirs. The three methods aim to enhance coarse grid simulation models to accurately capture the production behavior that occurs during the initial transient time. We apply these methods in unconventional reservoir simulation models of different complexity (one-dimensional explicit matrix-fracture, small scale Embedded Discreet Fracture Modeling (EDFM) and single-stage hydraulic fracture EDFM with several thousand matrix-fracture connections) and compare them with existing methods such as global or local grid refinements.
Three methods to account for transient effects in reservoir simulation for unconventional reservoirs are proposed. All methods modify the transmissibility between matrix-fracture connections over time. The first two methods include an analytical expression based on a single-phase pressure diffusivity model. These methods compute the new transmissibility over time based on the coarse grid model properties. We tested these methods using models with different complexity and compared them with refined models. The third method uses two artificial neural networks. We train the first network using the refined grid to predict fluid flow at the fracture face. The second network is trained as a reverse proxy on the coarse model to compute transmissibility from flow rate.
Our analytical modifications provide accurate results during the early times and transition smoothly to the pseudo steady state of late times, while using only a fraction of the grid cells and reducing CPU time up to two orders of magnitude compared with the refined models.
We test our machine learning method in a one-dimensional case and compare it with the refined and analytical model shown in this work. We observed that when the coarse model fits within the assumptions of the single-phase pressure diffusivity model, both approaches give close results. We show that when rock compaction is included, the machine learning approach can capture the additional physics and produces a more accurate production profile than the analytical-based modification.
We present three new transient transmissibility modifications to improve the accuracy of coarse resolution simulation of unconventional reservoirs at early times. The analytical modifications of transmissibility provide a good approximation for models that lie outside the assumption of the single-phase pressure diffusivity model and due to its simplicity to implement, could be used as an improvement to static transmissibility coarse models. Additionally, we propose a new workflow that uses artificial neural networks to link high-resolution behavior with coarse-resolution properties. Finally, these methods could be extended to other applications such as geothermal processes and DPDK models.