Prediction of Ceiling Temperature in Tunnel Fires Under Canyon‐Wind Conditions Using Ensemble Learning

To overcome the limitations of traditional physical models in predicting the ceiling temperature of mountain tunnel fires under complex canyon‐wind environments, this study proposes a data‐driven ensemble learning prediction framework. Based on 1:20 reduced‐scale tunnel fire experiments, a feature dataset comprising 54,709 samples was constructed to comparatively analyze the predictive performance of Random Forest, XGBoost, and LightGBM models. Furthermore, the SHAP method was introduced to elucidate the marginal contributions of factors driving multi‐dimensional fire evolution toward the prediction target, as well as their coupling mechanisms underlying fire development. The results indicate that: (1) The LightGBM model achieves an optimal balance among predictive accuracy, generalization capability, and computational cost. (2) In an independent test set involving unknown extreme boundary conditions, the LightGBM model demonstrates acceptable robustness with R 2 and RMSE values reaching 0.7476 and 10.03, respectively, which fulfills the basic requirements for intelligent early warning of tunnel fires in complex wind zones. (3) Interpretability analysis reveals that the fire‐source location acts as the core variable dominating the evolution of ceiling temperature under the investigated conditions. It also confirms that variations in canyon‐wind yaw angle and wind speed jointly affect the induced airflow within the tunnel, exhibiting both competitive and synergistic effects in driving smoke spread.