Prediction of thermal conductivity of natural rock materials using LLE-transformer-lightGBM model for geothermal energy applications

Abstract

This paper focuses on the critical issue of predicting rock thermal conductivity in geothermal development and underground engineering, proposing a hybrid prediction method combining LLE (Locally Linear Embedding), Transformer, and LightGBM. The model leverages LLE for feature dimensionality reduction, Transformer for deep feature learning, and LightGBM for efficient regression analysis, achieving accurate and efficient predictions under complex and limited data conditions. Granite is selected as the study subject, and a multivariate database incorporating chemical compositions, physical properties, and environmental factors is constructed. An innovative inverse decomposition method is proposed to quantify the importance of original features during dimensionality reduction and prediction, revealing the significant impact of key factors such as surface distance, SiO₂, and CaO on thermal conductivity . Experimental results show that the proposed model significantly outperforms traditional methods in prediction accuracy, error control, and generalization capability, achieving an R² of 0.952, RMSE of 0.086, and MAE of 0.0739. A comparative analysis further demonstrates that the LLE-Transformer-LightGBM model outperforms other machine learning approaches, including Support Vector Regression (R² = 0.821, RMSE = 0.141) and Back Propagation Neural Network (R² = 0.860, RMSE = 0.148). These results highlight the superior predictive accuracy and robustness of the proposed model in handling complex geological datasets. This study not only provides a scientific basis for geothermal resource development, high-level radioactive waste disposal , and thermal management in deep engineering but also introduces an innovative feature importance quantification approach, offering new insights for complex data modeling and decision optimization in geotechnical engineering .

Topics & Concepts