A hybrid SVR-BO model for predicting the soil thermal conductivity with uncertainty

Kai Qi Li, Zhen Yu Yin, Yong Liu

Research output: Journal article publicationJournal articleAcademic researchpeer-review

9 Citations (Scopus)

Abstract

This study proposes a generalised framework for developing a hybrid machine learning (ML) model that combines support vector regression (SVR) with hyperparameter optimisation to predict thermal conductivity (k) with uncertainty. The framework contains four phases: data pre-processing, determining the best-performing hybrid model, selecting the optimal input combination, and uncertainty implementation. A database containing 2197 data points is first compiled to train the ML model. Three hyperparameter optimisation algorithms are adopted to tune hyperparameters, and their performance is evaluated by model evaluation metrics. Results show that SVR with Bayesian optimisation (SVR-BO) is the best-performing model since it produces more accurate predictions for k than models that employ grid and random searches. Given the sample insufficiency issue encountered in practice, the SVR-BO models with 144 input combinations are analysed. The compassion among models under various input combinations indicates that incorporating temperature as an additional input can provide moderate improvement in the accuracy and generalisability of the hybrid model. Based on the comparison, a five-input model is selected as the best candidate to implement the uncertainty evaluation for k. Results demonstrate that the predicted k possesses higher reliability for denser datasets and shows promising potential for applications in k with uncertainty assessments.

Original languageEnglish
Pages (from-to)258-274
Number of pages17
JournalCanadian Geotechnical Journal
Volume61
Issue number2
DOIs
Publication statusPublished - Feb 2024

Keywords

  • Bayesian optimisation
  • bootstrap technique
  • optimal input combination
  • soil thermal conductivity
  • support vector regression (SVR)
  • uncertainty

ASJC Scopus subject areas

  • Civil and Structural Engineering

Fingerprint

Dive into the research topics of 'A hybrid SVR-BO model for predicting the soil thermal conductivity with uncertainty'. Together they form a unique fingerprint.

Cite this