ABSTRACT

Pillar-gridding concept has been used in the conventional simulators to represent geological heterogeneities. The model construction was initiated by grid-cells generation and geostatistics application to populate properties based on underlying grid. This raises short-correlation-length, which spurious artefact in modelling since length-scale of variability is controlled by grid-cells, not the geology. Many problems are mostly encountered in the initial stages of grid-based modelling, especially for large heterogeneous reservoirs which entails detailed model.

New modelling workflows have been developed in which surface-based approach is applied in constructing the model. The surface-based modelling creates volumes bounded by surfaces to represent heterogeneity without being controlled by the grid-cells. Main advantage of this approach is ability to represent complex reservoirs in more accurate way. This leads to the reduced running time, since the model is using fewer cells.

Surface-based model can be performed with short-correlation-length produced by geostatistics approach, or with longer-correlation-length produced by assigning constant petrophysical property to the volumes that define internally homogenous rock. This project will focus on examining the effect of eliminating any short-correlation-length heterogeneity. To achieve this, a quantitative assessment is conducted by comparing longer-correlation-length case against the well-known SPE 10th model. This study is commenced by running a base case with thousands of grid-cells. Binning process is then conducted to group the properties. Various bin numbers are developed to investigate how increasing correlation-lengths affect flow simulation. Furthermore, clustering is conducted with image filtering to group the bins into flow units; therefore, geological domain with longer-correlation-length but internally homogeneous created. Hence, clustering is eliminating small features.

Comparison shows, eliminating short-correlation-length doesn't matter in models, since longer-correlation-length case only directs to the result error of approximately 10%. Moreover, 3D visualization of property supports the fact that longer-correlation-length case is similar to the base.

This content is only available via PDF.
You can access this article if you purchase or spend a download.