Evaluating and predicting deliverability of natural gas storage sites using stacking machine learning models

Abstract

Underground natural gas storage (UNGS) is crucial for balancing energy supply and demand, and supporting renewable energy integration. This study evaluates the distribution and deliverability of UNGS sites across the United States from 2014 to 2024, focusing on depleted fields, salt domes, and aquifers. A stacking model that integrates Random Forest, XGBoost, Support Vector Regression, and K-Nearest Neighbors is proposed, with the input parameters including total field capacity, base gas capacity, working gas capacity, surface temperature, site location, and company ownership type. Results show that UNGS sites in the Northeast region have the highest number of storage facilities, whereas the South-Central region exhibits significant capacity and variability. Depleted fields are the most common sites and generally have relatively low median capacities but present the potential for high-capacity storage in specific fields. Interstate pipeline companies and independent operators demonstrate high median deliverability and great variability. The stacking model can achieve superior accuracy across all storage types and outperform individual models. The SHapley Additive exPlanations (SHAP) sensitivity analysis shows that working gas capacity is the dominant factor influencing deliverability, followed by location, base gas capacity, and total field capacity. These findings offer valuable insights for optimizing UNGS operations and guiding strategic energy decisions.