Study On The Data-driven Model Of Personal Thermal Comfort In Xi’an Office Building In Winte

In recent years,the global energy crisis has intensified.In the current office buildings often use centralized air conditioning system,which unreasonable control behavior will lead to the indoor environment too cold or too hot,which not only wastes a lot of energy and reduces people’s thermal comfort level.In order to balance the contradiction between reducing energy consumption and improving thermal comfort,personal thermal comfort systems are an effective solution.With the development of data science,it is possible to use machine learning to predict personnel thermal comfort.When using personalized thermal comfort devices to regulate the local microenvironment of people,it is necessary to establish a specific personal thermal comfort model.Among them,human skin temperature becomes the key physiological factor in the study of thermal comfort.In this paper,12 subjects were selected to participate in three different winter heating conditions(21～24℃,24～27℃,27～30℃).The experiment continuously collected human physiological parameters and environmental parameters,and collected each subject’s thermal sensation,thermal preference and thermal satisfaction vote,to fully understand the interaction between indoor subjects and the environment.In this paper,four algorithms in machine learning(support vector machine,decision tree,K-nearest neighbor and integrated algorithm)are used to initially establish the thermal comfort evaluation model of the subjects for the overall indoor environment,and five evaluation indicators are selected to measure the performance of the model.This includes accuracy,precision,recall,F1,and Receiver Operating Characteristic(ROC).In the process of training the model,the negative effect of class imbalance on the model performance is also found.In the experimental phase,the superior performance of the proposed method is demonstrated by comparing four different advanced machine learning models(support vector machine,decision tree,integrated algorithm and K-nearest neighbor).In the case of all variables as inputs,the proposed Thermal Sensation Vote(TSV)model predicted the actual accuracy of 95.8%.The personal classification model based on Bayesian optimization technology was used to adjust the hyperparameters of the machine learning algorithm.The accuracy of the optimized personal thermal comfort model was improved by 13.3%.A conditional adversarial neural network framework suitable for generating thermal comfort samples was constructed through experiments,and the generalization ability of thermal comfort prediction models established by various machine learning algorithms was further improved,with the highest accuracy increased by 13.4%.Therefore,conditional adversarial neural network technology is used to generate thermal comfort data close to the real value,which solves the problem of insufficient sample number and successfully improves the performance of the model.Moreover,the new data is used to verify the existing predicted thermal comfort model,and the accuracy reaches75%,which proves the feasibility of the thermal comfort prediction model proposed in this paper.The experimental results of this study show that the personal thermal comfort model of Xi ’an district heating condition based on machine learning can accurately and quickly realize the evaluation and classification of personal thermal comfort level,and can provide a valuable reference for the establishment of automatic control personal thermal comfort model.