A Novel Similarity-Based Upscaling Method for Highly Heterogeneous Reservoir Models

Abstract

The paper presents a novel method for upscaling geological model grids of hydrocarbon reservoirs, particularly well-suited to highly heterogeneous formations. To determine the optimal number of layers in the upscaled model, the Lorenz coefficient was used. A rapid decline in its value was interpreted as a significant loss of geological information, providing a quantitative criterion for limiting vertical coarsening. Hydraulic Units (HU) were calculated based on the analysis of Reservoir Quality Index (RQI) and Flow Zone Indicator (FZI) parameters. These units served as the basis for transforming the reference model into a rock-type model. Advanced image comparison techniques based on deep artificial neural networks were used to compute the similarity between adjacent layers. This enabled the identification and merging of geologically similar layers while preserving the reliability of the upscaled model. The proposed method outperforms conventional upscaling approaches in terms of both accuracy and computational efficiency. Notably, it eliminates the need for time-consuming optimisation procedures, which are commonly required in standard workflows. Furthermore, the developed algorithm is grounded in robust theoretical principles related to reservoir rock classification and allows continuous improvement of results via retraining or replacement of the image analysis module.