Development of bubble point pressure and oil formation volume factor models using pressure-volume-temperature data

Abstract

The Pressure-Volume-Temperature (PVT) properties of crude oil are typically determined through laboratory analysis during the early phases of exploration and field development. However, due to extensive data required, time-consuming nature, and high costs, laboratory methods are often not preferred. Machine learning, with its efficiency and rapid convergence, has emerged as a promising alternative for PVT properties estimation. This study employs the modified particle swarm optimization-based group method of data handling (PSO-GMDH) to develop predictive models for estimating both the oil formation volume factor (OFVF) and bubble point pressure ( P b ). Data from the Mpyo oil field in Uganda were used to create the models. The input parameters included solution gas-oil ratio ( R s ), oil American Petroleum Institute gravity ( API ), specific gravity ( SG ), and reservoir temperature ( T ). The results demonstrated that PSO-GMDH outperformed backpropagation neural networks (BPNN) and radial basis function neural networks (RBFNN), achieving higher correlation coefficients and lower prediction errors during training and testing. For OFVF prediction, PSO-GMDH yielded a correlation coefficient ( R ) of 0.9979 (training) and 0.9876 (testing), with corresponding root mean square error ( RMSE ) values of 0.0021 and 0.0099, and mean absolute error ( MAE ) values of 0.00055 and 0.00256, respectively. For P b prediction, R was 0.9994 (training) and 0.9876 (testing), with RMSE values of 6.08 and 8.26, and MAE values of 1.35 and 2.63. The study also revealed that R s significantly impacts OFVF and P b predictions compared to other input parameters. The models followed physical laws and remained stable, demonstrating that PSO-GMDH is a robust and efficient method for predicting OFVF and P b , offering a time and cost-effective alternative.