Fast History Matching and Robust Optimization Using a Novel Physics-Based Data-Driven Flow Network Model: An Application to a Steamflood Sector Model

Abstract

Summary Full-fidelity models can be computationally expensive during history matching (HM) and robust optimization as these problems typically require hundreds of simulations. Previously, we have implemented a physics-based data-driven flow network model, general purpose simulator-powered network model (GPSNet), that serves as a surrogate without the need to build the 3D full-fidelity model. In this paper, GPSNet is enhanced to GPSNet-2D to better capture thermal processes, especially gravity drainage. This enhancement successfully enables a rapid HM and robust optimization for a steamflood sector model. In GPSNet-2D, the reservoir is discretized into a series of 2D connections (x-z planes) between well completions. These connections capture the main areal/vertical flow paths, and the grid properties of each connection are interpreted with historical data. Steam segregation and heat loss are included to better represent the subsurface physics. When simulating GPSNet-2D using a commercial simulator, an equivalent 3D Cartesian model is designed where vertical slices correspond to the interwell connection planes. Thereafter, an iterative HM is conducted using the ensemble smoother with multiple data assimilation. The best-matched model is then used for steam injector control optimization. The GPSNet-2D model is first validated through a synthetic steamflood case. HM result shows that the GPSNet-2D model not only aligns closely to field-scale volumetric data but also yields good well-level matches, including bottomhole pressure (BHP)/bottomhole temperature (BHT) and phase rates. Then, GPSNet-2D is successfully applied to steamflood HM and robust optimization for a 36-well sector of a heavy oil field in the San Joaquin Valley in North America. Again, the calibrated GPSNet-2D model demonstrates its capability and reliability by generating accurate field-scale match results and reasonable matches for most of the wells. For robust well control optimization, we select nine realizations from the posterior ensemble and maximize NPV under well constraints simultaneously. The optimal case improves the average NPV by 27% over the reference case. The integration with a commercial simulator in GPSNet-2D provides flexibility to account for complex physics in the thermal recovery processes, such as steam segregation near injectors, fast steam breakthroughs at producers, and heat loss to the overburden/underburden formations. Unlike traditional simulation that relies on a detailed characterization of geological models, the GPSNet-2D model only requires well volumetric production/injection data and its approximate trajectory and can be generated and updated rapidly. In addition, GPSNet-2D also runs much faster (minutes) than a full-fidelity thermal model due to having much fewer gridblocks in the model. To our knowledge, this is the first time a physics-based data-driven network model integrated with a commercial simulator is demonstrated via a field steamflood case. Unlike approaches developed with analytic/empirical solutions or research simulators, the use of a commercial simulator makes it possible to extend flow network modeling to simulate enhanced oil recovery processes more realistically. It serves as an ideal surrogate model for both fast and reliable decision-making in reservoir management.