Abstract
Liquid dropout occurs in gas condensate reservoirs below the dew point pressure around near wellbore region as a result of depletion from production of such reservoirs. Forecasting production as well as optimizing future recoveries of gas condensate reservoirs are highly desirable. This is not possible to achieve without accurate determination of liquid dropout viscosity (μ_c) below the dew point. The focus of research in past decades has been on the development of accurate viscosity prediction models below the dew point pressure to ensure accurate condensate production forecast. Gas condensate production forecast and optimisation around this region and condition are complicated due to unique gas condensate behaviour that violates thermodynamic laws.
Current methods are based on correlation estimation, however the accuracy of these correlations are less than satisfactory, and root cause is due to the miscapturing of complex behaviour of gas condensate reservoir near the wellbore region. These motivated the consideration of modern numerical approaches such as the Least Square Support Vector Machine (LSSVM) and Artificial Neural Network (ANN) used in this paper. These methods are considered as more data behaviour oriented, with the capability of capturing the fluid complexity of gas condensate in such conditions.
In this study viscosity of condensate phase near the wellbore region was modelled using machine learning techniques including ANN and LSSVM. For this purpose, over 300 viscosity data sets were collected from published literature and experimental studies worldwide. This databank includes API gravity, reservoir temperature, solution gas to oil ratio (Rs), specific gas gravity, fluid compositions and reservoir pressure.
Six well known previously published viscosity correlations refined using least-square approach to match the experimental data. Qualitative and quantitative error analysis of developed LSSVM and ANN showed their performance superiority over refined literature correlations. The new proposed models can be embedded as an extra feature of commercial reservoir simulation packages for optimization and future recoveries of gas condensate reservoirs
Current methods are based on correlation estimation, however the accuracy of these correlations are less than satisfactory, and root cause is due to the miscapturing of complex behaviour of gas condensate reservoir near the wellbore region. These motivated the consideration of modern numerical approaches such as the Least Square Support Vector Machine (LSSVM) and Artificial Neural Network (ANN) used in this paper. These methods are considered as more data behaviour oriented, with the capability of capturing the fluid complexity of gas condensate in such conditions.
In this study viscosity of condensate phase near the wellbore region was modelled using machine learning techniques including ANN and LSSVM. For this purpose, over 300 viscosity data sets were collected from published literature and experimental studies worldwide. This databank includes API gravity, reservoir temperature, solution gas to oil ratio (Rs), specific gas gravity, fluid compositions and reservoir pressure.
Six well known previously published viscosity correlations refined using least-square approach to match the experimental data. Qualitative and quantitative error analysis of developed LSSVM and ANN showed their performance superiority over refined literature correlations. The new proposed models can be embedded as an extra feature of commercial reservoir simulation packages for optimization and future recoveries of gas condensate reservoirs
| Original language | English |
|---|---|
| Article number | 106604 |
| Number of pages | 14 |
| Journal | Journal of Petroleum Science and Engineering |
| Early online date | 14 Nov 2019 |
| DOIs | |
| Publication status | E-pub ahead of print - 14 Nov 2019 |