Contents lists available at ScienceDirect Building and Environment journal homepage: www.elsevier.com/locate/buildenv Comparison of different window behavior modeling approaches during transition season in Beijing, China Yixuan Wei a , Haowei Yu b , Song Pan c,d,e,∗ , Liang Xia a , Jingchao Xie c , Xinru Wang a , Jinshun Wu f,∗∗ , Weijie Zhang b , Qingping Li g a Research Centre for Fluids and Thermal Engineering, Department of Architectural and Built Environment, University of Nottingham Ningbo China, Ningbo, 315100, China b College of Energy and Environmental Engineering, Hebei University Of Engineering, Handan, 056038, China c Beijing Key Laboratory of Green Built Environment and Energy Efficient Technology, Beijing University of Technology, Beijing, 100124, China d Engineering Research Center of Digital Community, Ministry of Education, Beiing, 100124, China e Beijing Laboratory for Urban Mass Transit, Beiing, 100044, China f College of Architecture and Civil Engineering, North China Institute of Science & Technology, Hebei, 065201, China g Beijing Institute of Residential Building Design and Research Co., LTD, Beijing, 100005, China ARTICLE INFO Keywords: Window behavior Logistic regression Markov model Artificial neural network Office building ABSTRACT Window operation is an important occupant behavior, and has significant impacts on building energy con- sumption. Recently, various stochastic and non-stochastic models have been proposed, aiming to describe oc- cupant window behavior based on several influencing factors. However, most of the employed methods are logit regression and Markov chain techniques, and the application of machine learning to model occupants' window behavior is rarely investigated. In addition, most published studies referring to occupants' window behavior have been carried out within European countries, where the influence of outdoor air quality is rarely considered. This study compares different models of occupants’ window behavior, including models based on logistic regression, Markov processes, and an artificial neural network. An artificial neural network model is proposed to explore the application and optimization of an artificial neural network algorithm under a condition of having less samples. Moreover, the outdoor fine inhalable particles (PM 2.5 ) concentration is considered as an influencing factor for building a window opening model for office buildings during the transition season in China. From this work, it is generally concluded that the PM 2.5 concentration and outdoor humidity should be considered in the modeling of occupant window behavior in Beijing, China. In addition, more true estimations can be obtained from artificial neural network models than from logistic regression models and Markov models. This result demonstrates that the proposed artificial neural network yields a prediction model of office window states with higher accuracy and better interpretability of highly correlated factors as compared to logistic regression models and Markov models. The proposed approaches provide a new and detailed way for engineers and building operators to better understand occupant window behaviors and their impacts on energy use in office buildings. 1. Introduction In recent years, research regarding building energy has gradually focused on occupancy behaviors, because indoor occupancy is an im- portant factor directly influencing building energy use [1,2]. The need to integrate occupancy behavior into building energy use has brought more awareness to window operation behavior. Specifically, various stochastic models of window operation have been proposed, aiming to capture occupants’ interaction with windows based on several influencing factors. Logistical regression is the most popular stochastic model used for a change of window state [3,4]. For example, a logistical regression method was used to infer the probability of opening and closing a window based on 15 dwellings in Denmark [5]. The indoor CO 2 con- centration and the outdoor temperature were identified as the most important influencing variables in determining the probability of opening and closing windows, respectively. More elaborate models were created by D. Calì et al. [6] based on logistical regression, to https://doi.org/10.1016/j.buildenv.2019.04.040 Received 25 January 2019; Received in revised form 24 March 2019; Accepted 19 April 2019 ∗ Corresponding author. Beijing Key Laboratory of Green Built Environment and Energy Efficient Technology, Beijing University of Technology, Beijing, 100124, China. ∗∗ Corresponding author. E-mail addresses: pansong@bjut.edu.cn (S. Pan), wujinshun2005@163.com (J. Wu). Building and Environment 157 (2019) 1–15 Available online 20 April 2019 0360-1323/ © 2019 Elsevier Ltd. All rights reserved. T