Ensemble machine learning prediction of LNG tank thermal stratification and boil-off gas generation: an experimental study with uncertainty quantification
Liquefied natural gas (LNG) storage tanks are susceptible to thermal stratification, a phenomenon that triggers rollover events and uncontrolled boil-off gas (BOG) generation, posing significant safety and economic risks. Traditional computational fluid dynamics (CFD) approaches offer mechanistic insight but require prohibitive computational resources and precise boundary conditions that are difficult to obtain under industrial operating conditions. Empirical correlations, while efficient, fail to capture the nonlinear, transient coupling between thermal stratification and BOG dynamics. This study presents an integrated experimental and machine learning framework for real-time prediction of BOG rates and rollover risk in a 1.0 m diameter LNG storage tank. A comprehensive dataset comprising 300 operational records was acquired over 70 h. Seven machine learning models were evaluated, and an ensemble of Random Forest, Gradient Boosting, and Multi-Layer Perceptron achieved the highest accuracy ( R 2 = 0.90, RMSE = 0.11 kg/h, MAPE = 3.30%), outperforming both empirical Chato correlations ( R 2 = 0.62) and CFD-RANS simulations ( R 2 = 0.78). Permutation importance revealed that the thermal stratification index and liquid level are the dominant drivers. Bootstrap 95% prediction intervals provided rigorous uncertainty quantification. Feature ablation confirmed that removing the stratification index degrades R 2 by 0.312. Time-series forecasting yielded R 2 = 0.968, with Bland–Altman analysis confirming clinical-grade agreement. These findings demonstrate that interpretable, low-cost ensemble learning can serve as a viable digital-twin alternative to first-principles modeling for industrial LNG storage monitoring.