Machine Learning Models for CO Storage Estimation in Natural Gas Reservoirs

Accurate estimation of CO storage capacity in depleted hydrocarbon reservoirs is crucial for carbon capture and storage (CCS) planning. The paper develops machine-learning (ML) models to predict the CO storage potential of natural gas reservoirs using UK Continental Shelf (UKCS) data, calculated based on reservoir properties such as original oil in place, recovery factor, temperature, pressure, rock volume, and porosity, as well as in situ CO density. The research applies three tree-based ML models Random Forest (RF), Gradient Boosted Trees (GBT), and Extreme Gradient Boosting (XGB) to this dataset. The results show that all models can approximate traditional engineering calculations efficiently, with XGB achieving the best performance (highest R and lowest RMSE). Complicated calculations, including estimations of recoverable oil and gas, reservoir attributes (depth, temperature, pressure, etc.), and the physical characteristics of the retained fluids, may be necessary for reconnaissance-type assessments of CO2 storage resources of known oil and gas reservoirs. These reconnaissance estimates can inform CCS site screening more rapidly than detailed reservoir simulations.