Use of Clustering Techniques for Automated Lumping of Components in Compositional Models

Abstract

Compositional simulation run times grow significantly as we increase the number of components used to characterize our fluid (SPE 69575). Therefore, having usable and practical models requires that we minimize the number of components without sacrificing prediction accuracy. In this paper we validate a novel approach that automates the compositional lumping as a part of the simulator pre-processing and allows quick evaluation of the impact on results and run-times of reducing the number of components in actual simulation runs. Different clustering techniques such as K-means or Agglomerative are applied on five different compositions from the literature which typically would require compositional modelling (gas condensate to volatile oil). The performance of these lumped compositions obtained from clustering are compared with an exhaustive brute-search lumping and the original full composition. These comparisons are made by simulating classical CCE & DLE or CVD lab experiments. The results are quantitatively assessed for proximity to the full composition simulation. With the techniques already validated, a preprocessor is developed that allows the user to input a full composition and set the number of components to be used for the run. These heuristic clustering techniques provide excellent results with minimal time. Although brute-force search may occasionally deliver marginally better outcomes, it does so at immense computational costs and any advantage vanishes after regression. To the best of our knowledge, advanced clustering techniques have not previously been applied to the problem of lumping as the industry has relied mostly on theoretical or empirical arguments to prescribe the lumping approach, to be carried out manually, with the occasional study on brute force search (SPE-170912). An additional novelty is to automate compositional lumping in the simulator preprocessor, allowing for accelerated validation of the lumping approach under the expected reservoir conditions. The speed and flexibility of the approach makes it an excellent practical option to test and scale the number of components used in compositional models.