A physics-based data-driven simulation method for water flooding in heterogeneous reservoirs with large pores

Traditional grid-based reservoir numerical simulation methods face challenges, including low computational efficiency and poor convergence, when applied to full-field reservoir models. To address these limitations, data-driven and physics-based surrogate models have been developed as viable alternatives. Among these, physics-based flow network models that incorporate seepage mechanisms have demonstrated significant improvements in computational performance. However, existing flow network models often neglect critical dynamic factors, particularly in strongly heterogeneous reservoirs with large pores. Specifically, the time-varying physical properties of reservoir rocks and fluids, as well as the high-speed nonlinear seepage behavior in large pores during long-term water flooding, are not adequately considered, leading to limitations and inaccuracies in predictions. The study proposes an enhanced flow network model that integrates time-varying permeability, relative permeability, and viscosity relationships derived from long-term water-flooding experiments. Additionally, the model accounts for high-speed nonlinear flow in large pores (thief zones) using the Forchheimer equation, while maintaining Darcy flow in small pores. The reservoir is discretized into one-dimensional flow units with nodal-based pressure and saturation solutions computed using the Buckley–Leverett theory and Newtonian iteration. Furthermore, a Bayesian framework-based objective function is established for historical data assimilation, and the Simultaneous Perturbation Stochastic Approximation algorithm is employed to automate the history matching process. This approach enables the dynamic correction of key parameters, including permeability, relative permeability endpoints, and the non-Darcy flow coefficient. The results demonstrate that the proposed model effectively captures the spatial distribution and temporal evolution of reservoir properties and fluid dynamics under long-term water flooding. The model exhibits superior convergence in history matching compared to conventional methods while achieving a 20-fold improvement in computational efficiency over traditional full-scale grid-based simulations. This advancement provides a more robust and efficient tool for reservoir performance prediction and optimization in highly heterogeneous reservoirs.