Robust computational approach to determine condensate liquid viscosity

Abstract

The production of gas condensate from condensate gas reservoirs (GCR) presents significant challenges in reservoir engineering management, production, and operations. A crucial factor affecting the production and transport of condensate gas is the condensate liquid viscosity (μlc), which is vital for equations of state and for establishing relationships between PVT properties. This study aims to predict viscosity using five input variables: condensate gravity (API), initial gas-to-condensate ratio (RS), pressure (P), gas specific gravity (γg), and temperature (T). To achieve this, four robust models, previously unused in this domain, were used. Data from 2160 records were gathered from Iranian reservoirs, with 2114 data sets retained after outlier elimination using k-means clustering. The study combines multilayer perceptron (MLP) and distance-weighted k-nearest neighbor (DWKNN) networks with two optimizers, independent component analysis (ICA) and the gravitational search algorithm (GSA), to predict μlc. The results indicate that the AI-based hybrid model achieves significantly greater accuracy than the four empirical equations evaluated, with the DWKNN-GSA model outperforming the others in terms of accuracy (R² = 0.9998, RMSE = 0.0037 cP). Correlation analysis reveals that P, API, and RS highly influence μlc, whereas T and γg have a low impact. A heat map diagram further highlights that γg exerts the highest effect, while API has the lowest impact on μlc. The approach used in this study demonstrates significantly higher accuracy in predicting μlc compared to other published methods and could be applied to the prediction of similar parameters.