YAO Zhaoming, WANG Xun, QI Jian
Journal of Engineering Thermophysics.
2024, 45(5):
1440-1449.
The thermal conductivity of soil is a vital parameter in describing its heat transfer properties. Accurate prediction and sensitivity analysis of this parameter can aid in assessing the thermal response in geotechnical engineering and prevent deformation and damage in projects. Based on thermal conductivity experiments by Kersten’s team, we analyzed the factors influencing this parameter. we considered introducing a temperature variable into the traditional empirical formula and conducted validation, resulting in an improved formula with good applicability to clay. Using artificial intelligence algorithms, we established a prediction model for thermal conductivity. The model uses soil type, dry density, water content, and temperature as input variables. Our analysis showed that the Random Forest model, Radial Basis Function Neural Network (RBFNN), and Whale Optimization Algorithm Backpropagation Neural Network (WOA-BP) could all accurately predict thermal conductivity. Among these, the WOA-BP model demonstrated the best performance, followed by Random Forest and RBFNN.We tested the prediction model using a new sample set and found that the model still performed well, indicating a certain level of generalization ability. We employed a Monte Carlo simulation for parameter sensitivity analysis of the improved empirical formula. With the Random Forest model, we ranked feature importance to evaluate the impact of different input variables on model output. Finally, we calculated the sensitivity of influencing factors using a weighted product method combined with WOA-BP. The results from all three methods were consistent. They indicated that the sensitivity of thermal conductivity to changes decreases in the order of water content, dry density, temperature, and soil type.