Production rate forecasting using pressure and saturation estimates near the wellbore

Abstract

Production flow rate forecasting is crucial for effective reservoir management decisions. Usually, this forecasting relies on either full-scale reservoir models, which demand significant computational resources, or highly simplified surrogate models like the Capacitance–Resistance Model (CRM). While CRM offers computational efficiency, it faces considerable challenges when applied to multilayer reservoirs or scenarios with varying fluid saturations. We present a novel method for flow rate forecasting, occupying an intermediate position between these two approaches. The model consists of several components: separate machine learning models are used to predict pressure and fluid saturation in each grid cell perforated by the well. The predicted pressure and saturation values are then used to estimate well production rates. This workflow preserves the physical interpretability: outputs of intermediate models represent primary dynamic variables and the weighting coefficients of the flow rate estimation model can be interpreted as connection factors of the grid blocks. At the same time, it overcomes some of the limitations of both full-scale reservoir model and CRM-like surrogate models. Unlike CRM, each well is considered as a collection of cells for better representation of structural heterogeneity. Additionally, by considering fluid saturation as a dynamic property, it can manage problems with varying fluid phase composition. The focus only on values near the wellbore facilitates rapid forecasting compared to full-scale modeling. We test the proposed approach using two reservoir models: a simplistic synthetic reservoir and a model of a real-world reservoir, and compare it with the results obtained using a commercial reservoir simulator and CRM. We observe a better accuracy against the CRM and a reasonable approximation of full-scale simulation results.