Interpretable Machine Learning of Salt Cavern Volume Shrinkage: Implications for Long-Term Stability Engineering

The volume shrinkage of salt caverns is a key parameter for evaluating their long-term stability and operational lifespan. However, existing methods cannot establish a volume shrinkage prediction model that takes multiple factors into account. This study constructed a database of salt cavern volume shrinkage using 406 cases and established eight prediction models based on machine learning. The prediction curve and residual distribution plot indicate that the Gradient Boosting Machine (GBM) model is the optimal predictive model. Its R2 (correlation coefficient) and RMSE (root-mean-square error) are 0.880 and 4.795, respectively, with residuals distributed within the range of −7.5 to 7.7. Based on this, shapley additive explanation (SHAP) analysis was conducted, revealing that low operational pressure (LP), cavern depth (CD), high operational pressure (HP), creep parameters n and A are the most significant factors influencing the volume shrinkage. This study marks the first application of machine learning combined with SHAP analysis in the long-term stability analysis of salt caverns. The SHAP values for LP, CD, n, A, and HP are −4.47, −2.42, 1.31, 1.01, and −0.76, respectively. The absolute values of SHAP for all other variables are less than 0.3. The results quantitatively elucidate the complex mapping relationship between the volume shrinkage and input variables. The findings of this study can make substantial contributions to the structural design and protective management of salt cavern engineering during the operational process.