基于模型迁移学习的木材导热系数预测  

Model-based Transfer Learning for Prediction of Multi-factor Wood Thermal Conductivity

作　　者：冯彦皓 俞自涛[1,2] 陆江[3] 徐旭 FENG Yanhao;YU Zitao;LU Jiang;XU Xu(Institute of Thermal Science and Power Systems,Zhejiang University,Hangzhou 310027,China;State Key Laboratory of Clean Energy Utilization,Zhejiang University,Hangzhou 310027,China;School of Civil Engineering and Architecture,Zhejiang University of Science and Technology,Hangzhou 310023,China;College of Metrology and Measurement Engineering,China Jiliang University,Hangzhou 310018,China)

机构地区：[1]浙江大学热工与动力系统研究所,杭州310027 [2]能源高效清洁利用全国重点实验室(浙江大学),杭州310027 [3]浙江科技学院建筑工程学院,杭州310023 [4]中国计量大学计量测试工程学院,杭州310018

出　　处：《工程热物理学报》2024年第3期865-872,共8页Journal of Engineering Thermophysics

基　　金：国家自然科学基金资助项目(No.51978623)。

摘　　要：木材导热系数对木结构建筑能耗精准预测具有重要意义,然而目前文献中的模型难以提供全面而准确的预测工具,且一些机器学习模型所考虑的因素和样本数有限。本文构建1941年至今含多种特征的导热系数数据库,并开发对比神经网络和迁移学习模型的预测精度。结果表明,结合迁移学习后,导热系数预测精度可提升23%以上,且迁移学习性能在0~0.3 W·m^(-1)·K^(-1)上精度提升明显,更适合于木结构等传热应用中。本文通过创新的模型参数微调迁移策略能为小样本、含大量缺失值的数据集和简单网络结构的导热系数预测提供思路。Wood thermal conductivity is important for accurate prediction of energy consumption in timber structure;however,current models in the literature are difficult to provide comprehensive and accurate prediction tools,and some machine learning models have limited factors and sample sizes to consider.In this paper,a database of thermal conductivity coefficients containing multiple features from 1941 to the present was constructed,and a comparison of the prediction accuracy of neural networks and transfer learning models was developed.The results show that the prediction accuracy of thermal conductivity can be improved by more than 23% after combining transfer learning,and the transfer learning performance is significantly improved on 0~0.3 W·m^(-1)·K^(-1),which is more suitable for heat transfer applications such as timber structures.The novel transfer strategy with fine-tuning can provide ideas for thermal conductivity prediction of small samples,data sets with large number of missing values and simple network structures.

关 键 词：木材导热系数 迁移学习 神经网络 

分 类 号：TK124[动力工程及工程热物理—工程热物理] S781.37[动力工程及工程热物理—热能工程]