Toward Predictive Models for Estimation of Bubble-Point Pressure and Formation Volume Factor of Crude Oil Using an Intelligent Approach

Abstract

-Accurate estimation of reservoirs fluid properties, as vital tools of reservoir behavior simulation and reservoir economic investigations, seems to be necessary. In this study, two important properties of crude oil, bubble point pressure (P b ) and formation volume factor (B ob ), were modelled on the basis of a number of basic oil properties: temperature, gas solubility, oil API gravity and gas specific gravity. Genetic programming, as a powerful method, was implemented on a set of 137 crude oil data and acceptable correlations were achieved. In order to evaluate models, two test datasets (17 data for P b and 12 data for B ob ) were used. The squared correlation coefficient (R 2 ) and average absolute relative deviation (AARD %) over the total dataset (training + test) are 0.9675 and 8.22% for P b and 0.9436 and 2.004% for B ob , respectively. Simplicity and high accuracy are the advantages of the obtained models.

Data Analysis and Symbolic Regression Models for Predicting CO and NOx Emissions from Gas Turbines

“…Ref. [12] presents the formulae for estimating bubble-point pressure and the formation volume factor of crude oil with four basic oil properties: temperature, gas solubility, oil API gravity and gas-specific gravity as predictors. In [13], the SR methodology was used to develop a correlation to predict thermodynamic conditions for hydrates' formation.…”

Current Approaches in Prediction of PVT Properties of Reservoir Oils

“…In 2016 Iranian researchers D. Abooali and E. Khamehchi developed a method for the evaluation of PVT properties of crude oil such as bubble point pressure (Pb) and bubble point oil formation volume factor (Bob) using the genetic programming (GP) method [22]. The model based on parameters such as the reservoir temperature, specific gas gravity, solution gas oil ratio and specific oil gravity ( API  ).…”

Modeling of Catalytic CO2 Methanation Using Smart Computational Schemes

“…The GP approach was first introduced and developed by Koza 44 for mathematical and optimization targets. GP is now able to perform various operations similar to a machine learning methodology by producing randomly created mathematical functions in the form of “genes”, syntactic tree structures which are chromosome‐like and operate using input data 45, 46. Unlike conventional regression methods that require user‐specified models for fitting data, GP is implemented as a symbolic regression method meaning that the algorithms themselves search for both the model form and best fit.…”

“…Unlike conventional regression methods that require user‐specified models for fitting data, GP is implemented as a symbolic regression method meaning that the algorithms themselves search for both the model form and best fit. GP normally takes a multi‐gene approach rather than one, which normally results in simpler functions compared to monolithic GP models 46, 47.…”