Calculation of the Thermal Conductivities of Fine‐textured Soils Based on Multiple Linear Regression and Artificial Neural Networks

Haiyan Wen,Jun Bi,Ding Guo
DOI: https://doi.org/10.1111/ejss.12934
IF: 4.178
2020-01-01
European Journal of Soil Science
Abstract:The thermal conductivity of soils is an important parameter in environment, earth science and engineering applications. A new thermal conductivity model based on the Fredlund and Xing model was proposed to mimic the thermal conductivity of fine-textured soils from dryness to full saturation. However, the use of a measured thermal conductivity value at a certain saturation to calculate the parameter of the thermal conductivity model reduces the ease of use. Therefore, the objective of this study was to use multiple linear regression (MLR) and artificial neural networks (ANN) to improve the thermal conductivity model. The performances of the MLR-based model and ANN-based model were evaluated and compared with measured data and three existing empirical models at different saturation ranges. The results showed that the MLR-based model and ANN-based model performed best at low saturations, the ANN-based model performed best at saturations between 0.01 and 0.6, and the Johansen model showed the best fit to the measured data at saturations of above 0.6. In general, among the models, the ANN-based model performed best from dryness to full saturation (root mean square error (RMSE) = 0.098 Wm(-1) K-1 and average absolute deviation (AAD) = 0.071 Wm(-1) K-1), followed by the MLR-based model (RMSE = 0.110 Wm(-1) K-1 and AAD = 0.081 Wm(-1) K-1). The MLR-based model and ANN-based model are promising for the accurate calculation of the thermal conductivity of fine-textured soils from dryness to full saturation and can be incorporated into the numerical modelling of heat and mass transfers. Highlights The ANN-based model performed best among five models from dryness to full saturation, followed by the MLR-based model. The MLR-based model and ANN-based model showed the best fit to the measured data at low saturations. The ANN-based model showed the best fit to the measured data at saturations between 0.01 and 0.6. The Johansen (1975) model showed the best fit to the measured data at saturations of above 0.6.
What problem does this paper attempt to address?