Performance Evaluation of Machine Learning Algorithms in Predicting Dew Point Pressure of Gas Condensate Reservoirs

Accurate knowledge of the dew point pressure for a gas condensate reservoir is
necessary for optimizing mitigation operations during field development plan. This study explores the use of machine learning models in predicting the dew point pressure of gas condensate reservoirs. 535 experimental dew point pressure data points with maximum temperature and pressure of 304F and 10,500psi were used for this analysis. First, a standard multiple linear regression (MLR) was used as a benchmark for comparing the performance of the machine learning models. Artificial Neural Networks (ANN) [optimized for the number of neurons and hidden layers], Support Vector Machine (SVM) [using radial basis function kernel] and Decision Tree [Gradient boost Method (GBM) and XG Boost (XGB)] algorithms were then used in predicting the dew point pressure using gas composition, specific gravity, the molecular weight of the heavier component and compressibility factor as input parameters. The performances of these algorithms were analysed using root mean square error (RMSE), absolute average relative deviation percentage (AARD %) and coefficient of determination (R2). Besides the effect of the number of training data on the prediction performance of each algorithm, this work observes that neural network algorithm with single hidden layer and 4 to 5 neurons were better able to capture the intricate nature of gas condensate dewpoint pressures with least error margin.