A fast physics-based data-driven surrogate model for unconventional reservoirs with rapid decline and well interference

Abstract

Full-fidelity models are usually computationally prohibitive for unconventional fields, especially for simulating multiple wells with complex fracture networks. A physics-based data-driven surrogate, the general purpose simulator powered network model (GPSNet), was proposed previously for fast history matching and optimization in both waterflooded and steamflooded fields. GPSNet model serves as an ideal surrogate that can be created and updated rapidly without the need for detailed characterization of geological models. In this study, the GPSNet model is extended via a combination of two gridding systems to capture both the rapid decline in the early production stage and boundary effects and well interference in the late period for unconventional fields. A series of validation scenarios are carried out to demonstrate the necessity and effectiveness of the new design, followed by an application to a synthetic case with four multiple transverse fracture wells (MTFWs).GPSNet equivalently represents the entire reservoir via a network of 1D connections between well completions, which capture the primary flow paths in the reservoir. These connections are discretized into gridblocks, and their properties are adjusted to match historical production data. This adjustment takes into consideration factors, e.g., transmissibility, pore volume, saturation, capillary pressure, relative permeability, etc., to accurately depict the field conditions. To simulate this physics-based network model using a commercial simulator, an equivalent Cartesian model was created where each row represents one connection out of the flow network. Subsequently, the ensemble smoother with multiple data assimilation (ESMDA) algorithm is used to tune the model parameters during the history matching (HM) procedure. The resulting representative model from HM can then serve as a reliable surrogate for rapid production prediction, well control optimization, robust field development optimization, etc.To validate the new GPSNet model for unconventional reservoirs, various tests are initially performed using a synthetic case with a single-well two-wing-fracture system. The validation outcome unveils the necessity of employing two gridding systems and ensuring enough grid resolution. The HM results demonstrate that the predictions of the best-designed GPSNet model align closely with the volumetric observation data. After that, the GPSNet model is successfully applied to a challenging four-MTFW synthetic problem with rapid decline, boundary effects, and well interference. Once again, the calibrated GPSNet model showcases its capability and reliability by delivering accurate well-level matches for all the wells. What is more, it computes 120 times faster by employing only 0.06% of the gridblocks than the full-fidelity model.