Regional Assessment of Natural Gas Storage Deliverability in South-Central U.S. Using Stacking Machine Learning Models

Summary Underground natural gas storage (UNGS) is an important mechanism for stabilizing the supply and demand of natural gas, especially as more and more renewable energy sources are added to the power grid and the usage of natural gas varies with the seasons. In this study, a stacking machine learning framework is proposed to predict deliverability of UNGS sites in South-Central region, the United States. The stacking model integrates four base models—Random Forest, XGBoost, Support Vector Regression, and K-Nearest Neighbors—and utilizes six key factors: working gas volume, base gas volume, total storage capacity, surface temperature, geographic region, and operator type. The modeling shows excellent predictive performance, an R2 score of 0.9979, a RMSE of approximately 23,632 Mcf, and a MAE of 6,675 Mcf. SHapley Additive exPlanations (SHAP) analysis identifies working gas as the key predictor, with secondary factors varying with field type—base gas for salt domes and total capacity for depleted fields. These findings highlight the prospects of machine learning in enhancing the planning and operational efficiency of storage systems.