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 can experience thermal stratification, which leads to rollover events and uncontrolled boil-off gas generation. This creates safety and economic concerns for the industry. While CFD models provide insight, they demand high computational power and accurate boundary conditions that are hard to achieve in real-life scenarios.

Empirical correlations are quicker but cannot account for the complex, transient relationship between thermal stratification and BOG dynamics. This research proposes an experimental machine learning framework to predict BOG rates and rollover risk in a 1.0 m diameter LNG tank in real-time. The team gathered a dataset of 300 operational records over 70 hours.

By testing seven models, the ensemble of Random Forest, Gradient Boosting, and Multi-Layer Perceptron models showed the best performance with an R2 value of 0.90, RMSE of 0.11 kg/h, and MAPE of 3.30%. These results surpassed both the empirical Chato correlations (R2 = 0.62) and CFD-RANS simulations (R2 = 0.78). The study found that the thermal stratification index and liquid level were the most influential factors.

Bootstrap analysis provided a 95% prediction interval, while feature ablation showed that removing the stratification index decreased R2 by 0.312. Time-series forecasting reached an R2 of 0.968, with Bland–Altman analysis showing clinical-grade agreement between the model and actual data. The findings demonstrate that cost-effective ensemble learning can be a practical alternative to first-principles modeling for monitoring LNG storage systems.

This study was funded by the Shandong Engineering Research Center for Efficient Development of Oil & Gas Reservoirs and Geology-Engineering Integration, China.