当前位置: X-MOL 学术Energy Build. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
An ensemble strategy for transfer learning based human thermal comfort prediction: Field experimental study
Energy and Buildings ( IF 6.6 ) Pub Date : 2025-01-21 , DOI: 10.1016/j.enbuild.2025.115344
Kangji Li, Lei Chen, Yanpei Luo, Xiaotian He

Accurate prediction of human thermal comfort plays an important role in improving indoor comfort and reducing energy consumption. With the continuous innovation and improvement of physiological sensing technologies, data-driven models show better predictive performance than traditional predictive mean vote (PMV) model. However, the high performance of data-driven models requires sufficient and high-quality labeled data as input, which is difficult to access in this field. Transfer learning (TL) predicts target tasks by learning knowledge from similar domain, thus reducing the required data resource from target task. At present, transfer learning based thermal comfort predictions mostly applies model-based TL (MBTL) models and used public dataset as source domain. However, the predictive performance of such strategy lacks sufficient comparison and analysis. This study constructs three source domain datasets through two field experiments and one public data source, respectively. Their auxiliary training effects in TL models are fully compared. Three major TL methods with different mechanisms are implemented and their predictive performances are investigated. On this basis, a parallel ensemble strategy using hard / soft voting methods is proposed to improve the prediction accuracy and generalization performances simultaneously. Results show that: 1) if field experiment is available, thermal stepped experiment could provide more effective source domain data than public data source; 2) in addition to the applied MBTL model, the other two TL models (instance based TL, feature based TL) for this study are also very competitive, especially when target domain data is sparse. By hard voting method, the ensemble strategy containing these three sub TL models always achieve the best prediction accuracy under different amounts of target domain sampling data. Thus, it shows great potential in future practice of thermal comfort prediction.

中文翻译:


一种基于迁移学习的人体热舒适预测集成策略:现场实验研究



准确预测人体热舒适度对提高室内舒适度和降低能耗起着重要作用。随着生理传感技术的不断创新和改进,数据驱动模型显示出比传统预测平均投票 (PMV) 模型更好的预测性能。然而,数据驱动模型的高性能需要足够和高质量的标记数据作为输入,这在该领域很难获得。迁移学习 (TL) 通过学习相似领域的知识来预测目标任务,从而减少目标任务所需的数据资源。目前,基于迁移学习的热舒适预测主要采用基于模型的 TL (MBTL) 模型,并使用公共数据集作为源域。然而,这种策略的预测性能缺乏足够的比较和分析。本研究分别通过 2 个现场实验和 1 个公共数据源构建了 3 个源域数据集。对它们在 TL 模型中的辅助训练效果进行了充分比较。实施了三种具有不同机制的主要 TL 方法,并研究了它们的预测性能。在此基础上,提出了一种使用硬/软投票方法的并行集成策略,以同时提高预测准确性和泛化性能。结果表明:1)在田间实验可用的情况下,热阶梯实验可以提供比公共数据源更有效的源域数据;2) 除了应用的 MBTL 模型外,本研究的其他两个 TL 模型 (基于实例的 TL、基于特征的 TL) 也非常有竞争力,尤其是在目标域数据稀疏的情况下。 通过硬投票法,包含这三个 sub TL 模型的集合策略在不同数量的目标域采样数据下总是能达到最佳的预测精度。因此,它在未来热舒适预测的实践中显示出巨大的潜力。
更新日期:2025-01-21
down
wechat
bug