Data-driven modeling to predict the rate of Boil-off Gas (BOG) generation in an industrial LNG storage tank

Natural gas is stored as Liquefied Natural Gas (LNG) under cryogenic conditions in a storage tank. The tank is highly insulated; nevertheless, there is heat ingress from the surrounding, causing the generation of boil-off-gas (BOG). BOG generation over-pressurizes the tank and can lead to tank failure, so proper BOG management is critical to plant safety. This paper seeks to develop a data-driven model for a real industrial LNG terminal. Modeling was performed on historical real-time LNG data using different machine learning algorithms - Linear regression, Random Forest, and XGBoost. The performance of algorithms is analyzed based on R2, Mean-Absolute-Error (MAE %), and Root-mean-square error (RMSE %) values. The critical input features are calculated based on the Shapley additive explanation method (SHAP value); less important features were removed to decrease the model complexity. Our studies show that Random Forest outperforms the other two algorithms in terms of accuracy. The developed model can help plant operators make decisions quickly with better confidence.