Abstract: The concentration of harmful gases in underground caverns is closely related to excavation schemes and geological conditions; its accurate prediction under construction-geological impacts is crucial to construction safety management. However, it is challenging to extract useful information from harmful gas monitoring data due to the nonlinear coupling between features such as blasting parameters, bedrock types, and gas concentrations. This study presents an intelligent prediction method for harmful gas concentrations using integrated learning, incorporating the theory of SHapley Additive exPlanations (SHAP). The method conducts feature preprocessing using Principal Component Analysis (PCA), and uses the Tree-structured Parzen Estimator (TPE) algorithm to iteratively seek the optimal hyperparameters for the CatBoost (Categorical Boosting) model that is used to predict harmful gas concentrations. The SHAP explanatory framework is introduced to identify key factors affecting gas emission concentrations. Application in a case study of the diversion tunnel at the Xulong hydropower station shows that compared with the models of CatBoost, TPE-XGBoost, and TPE-LightGBM, the TPE-CatBoost model reduces the root mean square error by 48.9%, 40.2%, and 36.8% respectively, demonstrating a higher prediction accuracy. Integrating the SHAP theory reveals that PM10 and PM2.5 concentrations are more closely associated with blasting schemes, while CO and CO2 concentrations are more influenced by geological conditions such as groundwater state.

Key words: underground cavity, harmful gas, CatBoost integrated learning, TPE algorithm, SHAP theory