寒区隧道冻结特性及冻结深度预测

Freezing Characteristics and Freezing Depth Prediction of Tunnels in Cold Regions

为研究寒区隧道冻结特性并对冻结深度进行预测,采用数值模拟分析的方法,探讨兴安岭隧道温度场及冻结深度分布规律,并基于正交试验分析冻结深度各影响因素的敏感性及显著性。此外,提出一种基于XGBoost(extreme gradient boosting)和LightGBM(light gradient boosting machine)的混合预测模型对冻结深度进行预测。结果表明:1)距隧道洞口越远,围岩冻结深度越小;距隧道洞口越近,围岩冻结深度变化越大,在距洞口300 m之后,围岩冻结深度趋于稳定。2)各个断面不同部位冻结深度的变化规律基本一致,除仰拱部位冻结深度明显较大以外,其余各部位均无明显差别。3)各影响因素对冻结深度的影响敏感性排序为初始地温>最冷月平均气温>围岩比热容>围岩导热系数>衬砌导热系数>衬砌比热容。4)与传统单一模型预测方法相比,提出的基于XGBoost和LightGBM的混合预测模型精度较高,具有较强的适用性。

The on-site measured data of the Xing′anling tunnel and the numerical simulation results are studied together to investigate the characteristics of the tunnels′freezing depth in cold regions and further predict the freezing depth.Therefore,the temperature field and the distribution law of the freezing depth are explored.Subsequently,based on the orthogonal test,the sensitivity and significance of various influencing factors of the freezing depth are analyzed.Additionally,a hybrid prediction model is proposed based on extreme gradient boosting and light gradient boosting machine models to predict the freezing depth.The results reveal the following:(1)As the distance from the tunnel portal increases,the freezing depth of the surrounding rock shortens;moreover,the freezing depth of the surrounding rock varies sharply when close to the tunnel portal,whereas the freezing depth tends to be stable when the distance from tunnel portal is larger than 300 m.(2)The freezing depth′s change rule in different locations of each cross-section is similar.There is an obvious increase in the freezing depth in the inverted arch and no obvious difference in other parts.(3)The sensitivities of the freezing depth to various factors are ranked as follows:initial ground temperature>average temperature in the coldest month>specific heat capacity of the surrounding rock>thermal conductivity of the surrounding rock>thermal conductivity of the lining>specific heat capacity of the lining.(4)The proposed hybrid prediction model based on XGBoost and LightGBM model has a higher accuracy and stronger applicability than the traditional singlemodel prediction method.