Research on experimental tests and prediction models of thermal conductivity of freezing-thawing soil in the Kunlun Mountains

In order to explore the basic laws of freezing and thawing soil in the Qinghai-Tibet Engineering Corridor in the Kunlun Mountains area, the coefficient of thermal conductivity of 349 groups of drilling frozen soil samples and 245 groups of thawing soil samples is tested by the transient plane heat source method. The characteristics of five kinds of soil thermal conductivity distribution and natural moisture content, dry density and the partial correlation coefficient of thermal conductivity are analyzed, and the experience for both variables in fitting formula, support vector regression (SVR) and radial basis (RBF) neural network prediction model of thermal conductivity are established. The results show that the thermal conductivity of freezing-thawing soil is larger than that of fine-grained soil, and the thermal conductivity of freezing-thawing soil varies with the distribution of soil properties. Natural moisture content and dry density are positively correlated with thermal conductivity, and the partial correlation results of different soil types are significantly different. The binary empirical regression equation of typical soil thermal conductivity is shown as a nonlinear fitting result. The results of thermal conductivity prediction of the typical soil and freezing-thawing soil under three prediction models show that the prediction effect of fully weathered phyllite, breccia and gravel sand is better, and the prediction accuracy of SVR and RBF neural network of silty soil is also better. On the whole, the prediction effect of thermal conductivity on thawed soil is slightly better than that of frozen soil, and the prediction accuracy of thermal conductivity of breccias, silty soil and fully weathered meltwater is higher under the SVR and RBF neural network models. The prediction results and error analysis of the three thermal conductivity models show that the prediction results of the SVR and RBF neural network models are significantly better than that of the empirical fitting equation method. The prediction effect of SVR and RBF neural network prediction models varies with different soil thermal conductivities, and the overall prediction effect is similar, with higher prediction accuracy and wider range of soil application.