Predictive Modelling of Carbon Dioxide Injectivity Using SVR-Hybrid

Abstract

Southeast Asia is increasingly gaining attention as a promising geological site for permanent CO2 sequestration in deep saline aquifers. During CO2 injection into saline reservoirs, the reaction between injected CO2, the resident formation brine, and the reservoir rock could cause injectivity change due to salt precipitation, mineral dissolution, and fine particles migration. The underlying mechanisms have been extensively studied, both experimentally and numerically and the governing parameters have been identified and studied. However, the current models that have been widely adopted to investigate reactive transport and its impact on CO2 injectivity have fundamental limitations when applied to solve small, high dimensional, and non-linear data. The objective of this study is to develop efficient and robust predictive models using support vector regression (SVR) integrated with hyperparameter tuning optimization algorithms, including genetic algorithm (GA). To develop the model, 44 datasets are used to predict the CO2 injectivity change with its influencing variables such as brine salinity, injection flow rate, particle size, and particle concentration. The performance for each model is analyzed and compared with previous models by determination of coefficient (R2), adjusted determination of coefficient (R¯2), average absolute percentage error (AAPE), root mean square error (RMSE) and mean absolute error (MAE). The model with the highest R2 is selected as the predictive model for CO2 injectivity impairment during CO2 sequestration in a saline aquifer. The results revealed that both SVR and GA-SVR are able to capture the precise correlation between measured and predicted data. However, the GA-SVR model slightly outperformed the SVR model by a higher R2 value of 0.9923 compared to SVR with R2 value of 0.9918. Based on SHAP value analysis, brine salinity had the highest impact on CO2 injectivity change, followed by injection flow rate, particle concentration, and jamming ratio. It was also found that hybridization of genetic algorithm with support vector regression does improve the model performance contrary to single algorithm and contributes to the determination of the most impactful factors that induce CO2 injectivity change. The proposed model can be upscaled and integrated into field-scale models to improve the optimization of CO2 injectivity in deep saline reservoirs.