A Graph Network Based Approach for Reservoir Modeling

Reservoir management routinely requires assimilating historical data and predicting field performance against multiple production strategies before implementing them in the field. However, traditional numerical methods are often cumbersome to characterize, build and calibrate at a timescale that can be used reliably for such short-term decision cycles such as production forecasting, IOR optimization and production rate control. Simpler analytical models make assumptions and lack the rigor needed to adequately model these systems. Pure data-driven methods may lack physical insights or have limited range of applicability. Model fidelity, speed, interpretability, suitability to automate and ease-of-use are some key modeling traits that are desired for reservoir management purposes. In this work, we propose to use a reservoir graph-network modeling approach (RGNet), based on the concept of diffusive time of flight, to forecast well performance using routinely measured field measurements (e.g. bottomhole pressure and rates). We propose a novel, model order reduction method based on discretized time of flight for multiple wells with interference. It simplifies the 3D reservoir flow problem into a flow network representation that can be solved as a 2D simulation model with any general-purpose reservoir simulator. Parameters in RGNet model cover well productivity index, grid pore volume and transmissibility, which are estimated through a history-matching process. After history matching, multiple posterior RGNet models are generated to quantify subsurface uncertainties. The RGNet modeling approach allows various fluid-flow physics to be modeled within the grids and boundary conditions, and is applicable to a range of conventional and unconventional reservoirs with different flow mechanisms. We applied the proposed approach on a field case reservoir models for multiple wells with interference. By virtue of the reduced complexity, the modeling methodology is highly scalable and still retains physical interpretability. The calibration method produces parsimonious models and provides uncertainty estimates in history matching parameters with range of outcomes. In addition, the RGNet models are much faster to simulate, over 1000x speed up, compared with full-physics models. We then used RGNet models for well-control and flood optimization and achieved significant improvement over field net-present-values. Parameterization of the proposed reservoir graph-network modeling approach provides a unique and sustainable way to reduce model complexity needed for reservoir management purposes. The method is rooted in physical principles and provides an explainable dynamic reservoir model that can be effectively used to understand reservoir behavior and optimize performance. The lightweight model lends itself naturally to fast computation that are required for scenario analysis and optimization.