Analysis and modelling of gas relative permeability in reservoir by hybrid KELM methods
Zhang, Ning,Xi, Bin,Yu, Zhi,Fissha, Yewuhalashet,Taiwo, Blessing Olamide,Feng, Haibo
DOI: https://doi.org/10.1007/s12145-024-01326-2
2024-05-31
Earth Science Informatics
Abstract:Petroleum reservoirs are often influenced by various flow behaviours including the mixture of gas, water and oil. The gas relative permeability is used to estimate how much of the gas in the reservoir is producible at a given water saturation level. Therefore, the gas relative permeability is a significant parameter to characterize the behaviour of petroleum reservoirs. However, the measurement of gas relative permeability by traditional methods tends to be comparatively expensive and time-consuming. In the recent years, the machine learning techniques provided new alternatives for predicting the gas relative permeability. For this purpose, five new methods were proposed based on kernel extreme learning machine (KELM) technique. Five meta-heuristic algorithms were adopted to tune the model hyper-parameters of KELMs, i.e., butterfly optimization algorithm (BOA), tunicate swarm algorithm (TSA), Multi-verse optimizer (MVO), Golden jackal optimization (GJO) and Harris hawk's optimization (HHO). Five-fold cross validation was used to increase the model generalization. An extensive dataset from the experiments which contain 1024 data were taken to develop models. Four classical statistical indicators were used to measure the model performance, i.e., root mean squared error (RMSE), coefficient of determination (R 2 ), variance accounted for (VAF) and mean absolute error (MAE). In addition, two comprehensive manners, overall evaluation index (GI) and Taylor Diagram, were evaluated to provide overall model assessments. Proposed hybrid KELM models performed better than several other machine learning techniques. BOA-KELM model with swarm size 150 generated the best generalization for the testing set and could be recommended to predict the gas relative permeability with the same inputs used in this study. The detailed performance of BOA-KELM includes: training set (GI:0.1736; R 2 : 0.9902; RMSE: 0.7477; VAF: 99.0218; MAE: 10.6636), testing set (GI:0.4164; R 2 : 0.9789; RMSE: 0.5314; VAF: 97.8917; MAE: 4.1706). The mutual information technique was employed to examine the influence of influential factors to the model interpretation and it can be found that the gas saturation had a larger influence on the hybrid KELM models. When it was used as an individual input, the overall prediction decreased but acceptable prediction performance still can be obtained by hybrid KELM models. In the case of the gas saturation to be the only input, the best testing R 2 (0.94) could be generated by MVO-KELM which is higher than the R 2 from the empirical method named Corey-Brooks model and several other machine learning techniques. The main novelty of this study is that five new machine learning methods were proposed to predict the gas relative permeability and performed better than other empirical or machine learning techniques.
geosciences, multidisciplinary,computer science, interdisciplinary applications