【摘要】 热舒适度是衡量室内环境质量和影响人类健康的重要指标之一，是建筑、空调控制等系统智能化的重要参考依据，同时能够有效降低建筑热环境控制的能源需求．目前可穿戴设备如智能手环、柔性传感器等已广泛应用，可构建人体的健康大数据．但由于存在个体差异因素，不同个体对相同热环境所表现的生理热反应不同，基于单一个人的热舒适模型难以对群体热状态实现有效地预测．考虑到以往研究样本量相对较小、模型复杂难以部署等局限性，本文建立人工气候室，利用环境传感器和可穿戴设备收集了60名受试者的热舒适数据，采用机器学习实现人体热舒适度建模与预测．研究考虑身高、体重、性别等个体差异因素，采用XGBoost、随机森林和SVC共3种机器学习算法，得到了基于人体生理参数的增强型预测热态模型并对热舒适度进行分类．结果表明：对皮肤温度及其梯度进行归一化处理发现，归一化过程能够将冷不舒适、舒适、热不舒适3种状态拉开，有利于SVC算法在高维空间寻找最优超平面，对特征进行分类．对比归一化前后随机森林模型的特征重要性发现，归一化过程降低了体重、身高、性别等个体差异对模型预测效果的影响程度；在XGBoost、随机森林和SVC这3种机器学习算法中，SVC在测试集上的准确率和3种热状态的AUC值都高于XGBoost和随机森林，其分类效果和泛化能力最好．更多还原

【Abstract】 Thermal comfort is an important indicator of indoor environment quality and affects human health. It is an important reference for intelligent building systems,air conditioning control,and other systems. Moreover,it can effectively reduce the energy demand for controlling the thermal environment in buildings. Currently,wearable devices such as smart watches and flexible sensors are extensively used to compile comprehensive data on human health.However,due to individual differences,physiological thermal responses to identical thermal conditions vary,and it is difficult to effectively predict the group thermal state for a personal thermal comfort model. Considering the limitations of the relatively small sample sizes and complex model deployments in previous studies,this work established an artificial climate chamber with environmental sensors and wearable devices to collect thermal comfort data of 60subjects and leveraged machine learning to realize human thermal comfort modeling and prediction. Considering individual differences such as height,weight,and gender,three machine-learning algorithms,i.e.,extreme gradient boosting(XGBoost),random forest,and support vector classifier(SVC),were used to obtain an enhanced predictive thermal state model based on human physiological parameters and to classify thermal comfort. The results showed that the skin-temperature normalization process and its gradient result in three states(cold discomfort,comfort,and thermal discomfort). This facilitates the SVC algorithm to find the optimal hyperplane in high-dimensional space and classify the features. Comparative analysis of the feature importance of the random forest model before and after skintemperature normalization revealed that normalization reduces the influence of individual differences such as weight,height,and gender on the predictive effect of the model. Of the three machine learning algorithms,the accuracy of SVC on the test set and the area under the curve(AUC) values of the three thermal states were higher than those of XGBoost and random forest. Hence,SVC has the best classification effect and generalization capability.更多还原