MOBAPY: Modelling Household Adaptation to Summer Heat in Urban Environments Using Agent-Based Simulation

Abstract

With heatwaves becoming increasingly frequent and intense in recent years, summer comfort is now a major issue for the design of low-energy buildings. Currently, most French homes are not equipped with air conditioning. To prevent the widespread adoption of this energy-intensive equipment while limiting indoor thermal discomfort during heatwaves, it is crucial to accurately design and assess climate adaptation measures. Occupants actively employ various strategies to limit overheating inside their dwellings, adapting their actions on the specific constraints they face. However, they are still frequently treated as passive entities in building performance simulations. This paper introduces an agent-based model, MOBAPY, developed to simulate the summer adaptive behaviour of households in urban dwellings. Grounded in qualitative data derived from semi-structured interviews, the model integrates commonly reported actions for one occupant profile, including the operation of windows, solar shading devices, air conditioning, and fans, alongside clothing adjustments. MOBAPY is applied to one case study: an urban dwelling, simulated under multiple climate scenarios and varying behavioural constraints within a building energy modelling framework. The results highlight the significant impact of occupant behaviour on summer comfort results for this occupant profile in a well-insulated dwelling. Notably, under the most severe future climate scenario, the percentage of uncomfortable occupied time drops to 20% when behaviour is unconstrained, whereas it surges to 68% under highly constrained conditions.

1. Introduction

With climate change, extreme events such as heatwaves are expected to become more intense and more frequent. This will lead to an increased risk of overheating in both new and existing residential buildings, particularly in urban environments due to the urban heat island effect [1]. This risk can be reduced without relying on energy-intensive cooling systems, through passive cooling strategies such as solar shading and natural ventilation via window opening. However, the effectiveness of these strategies is strongly influenced by occupant behaviour (e.g., use of windows and shading devices) and the immediate environmental context.

Occupant behaviour (OB) significantly influences both the indoor environment and the energy consumption of buildings, through actions involving windows, solar shading, fans, thermostats and other devices [2]. It remains a major source of uncertainty in building performance simulation (BPS), mainly due to oversimplified assumptions [3]. In addition to emitting metabolic heat and pollutants due to their presence [4], occupants interact with their indoor environment to meet their needs and maintain comfort. However, in the building design process, occupants are still too often considered as passive elements, with limited consideration given to their adaptive behaviours in BPS [5]. The assessment of indoor comfort using a BPS should also be performed carefully by modellers. Indeed, thermal comfort indicators may not accurately reflect the variability of conditions within a space (spatial definition). This problem is exacerbated near glazed surfaces, due to simplified approaches used to determine the mean radiant temperature [6].

In order to simulate building performance while accounting for OB, it is necessary to model these behaviours and integrate them into BPS tools. Developing realistic OB models requires collecting data on occupant behaviour [7], through in situ measurements, laboratory experiments, surveys or even virtual reality studies [8]. Based on the quantitative or qualitative data collected, different types of models can be developed. These models are generally grouped into four categories: schedule-based and deterministic models, stochastic models, data-driven models and agent-based models (ABMs) [9].

Schedule-based and deterministic models are often considered too simple to represent the complexity of human behaviour. Stochastic and data-driven models rely solely on measurements without accounting for the reasons behind the behaviours. As a result, they are applicable only to the context in which the data was collected.

ABMs allow the behaviour of individual agents to be taken into account dynamically. An ABM consists of three key elements [10]: agents (autonomous entities with their own state), a topology (a set of rules governing relationships between agents) and an environment (in which agents interact with each other). ABMs are time-dependent [11], allowing the modelling of dynamic behaviour. Each agent reacts according to its perception of its environment and evolves within it [12] according to its own behavioural rules [13]. As they operate in a shared environment [14], agents may face constraints in their actions due to the presence of other agents with potentially conflicting needs. Agents can also be non-human, such as buildings [15].

ABMs offer multiple advantages: interactions between agents and with their environment can be explicitly represented; learning processes can be modelled; agent heterogeneity can be introduced; modelling is flexible; and the system represented can be complex and dynamic [12]. Many factors influencing occupant behaviour, including psychological or sociological dimensions, can also be taken into account [3].

The limitations of agent-based modelling are primarily related to its complexity, which stems from the complexity of the system being modelled, involving heterogeneous agents interacting dynamically within a variable environment. Simulation runtimes are often long, and a solid understanding of ABM principles is required [3]. Moreover, the model requires a large amount of input data [4], and assumptions must often be based on either observations or decision theories [13]. Coupling or communication with BPS tools can also be challenging, as it needs to be tailored to each simulation environment [3].

Table 1. Studies coupling ABM and BPS software.

Reference	ABM	BPS	Modelled Actions	Coupling Method
Vellei et al. [16]	Python	DIMOSIM	HVAC, Clothes	Not available
Moujalled [17]	Matlab 7.x (AdOCC)	TRNSYS 16	Windows, HVAC, Fan, Clothes	Directly in Matlab
Langevin et al. [18]	Matlab (HABIT)	EnergyPlus	Windows, HVAC, Fan, Clothes	BCVTB
Lee and Malkawi [19]	Matlab	EnergyPlus v7.1	Occupancy, Blinds, Windows, Doors, HVAC, Fan, Clothes	BCVTB and MLE+
Chapman et al. [20]	No-Mass	EnergyPlus	Occupancy, Artificial light, Blinds, Windows, HVAC, Clothes	FMI
Darakdjian [21]	No-Mass	EnergyPlus v8.2	Occupancy, Artificial light, Blinds, Windows, HVAC	FMI
Kim et al. [22]	NetLogo	EnergyPlus	Occupancy, Artificial light, Windows, Doors, HVAC	BCVTB
Putra et al. [23]	NetLogo	EnergyPlus	Artificial light, Blinds, HVAC, Fan, Clothes	Not available
Berger et al. [24]	NetLogo	EnergyPlus	Occupancy, Artificial light, Blinds, Windows, HVAC, Clothes	BCVTB/Python
Jia et al. [25,26]	PMFserv	EnergyPlus	Blinds, Windows, Doors	Manually then BCVTB
Alfakara and Croxford [27]	Repast	EnergyPlus	Occupancy, Windows, HVAC, Fan	FMI
Alfakara and Croxford [10]	Repast	Tas	Windows, HVAC	Not available
Gaaloul et al. [28]	BRAHMS	Matlab/Simulink	Occupancy, Appliances, HVAC	ICAr
Kashif [29]	BRAHMS	Matlab/Simulink	Occupancy, Appliances, Windows, HVAC, Clothes	Not available
Plessis et al. [30] Schumann et al. [31]	SMACH	BuildSysPro Modelica	Occupancy, Appliances, Blinds, Windows, HVAC	FMI

summarizes the studies that have coupled ABMs and BPS, indicating the occupant actions modelled and the coupling method used.

Several agent-based models have been developed for building application using programming languages such as Matlab [17,18,19] or Python [16]. More general ABM platforms have also been applied to the building sector, including Repast [10,27], PMFserv [25,26,32], NetLogo [22,23,24,33] and BRAHMS [28,29]. Additionally, several ABM platforms have been developed specifically for building applications, such as Occupancy Simulator [14], SMACH [30,31], No-Mass [20,21], BOT-ABM [11], AdOCC [17], Li-BIM [34] or HABIT [18].

Considering the impact of occupants’ behaviour on the indoor environment requires the integration of human activities into physics simulations. This is particularly complex as it first requires collecting data and to integrate it in a realistic way. Data collection and OB modelling requires an interdisciplinary collaboration between social sciences and building physics. Approaches from the human and social sciences are valuable for OB modelling: there are quantitative approaches with psychological and economic theories where behaviours are based on rational logic and planning [35], as well as qualitative approaches in sociology, geography and anthropology [36,37,38]. This raises the question of how to incorporate occupants’ statements (qualitative data) into BPS (quantitative models).

Several studies have proposed agent-based models for buildings, primarily addressing energy consumption. However, they often neglect occupants’ thermal management strategies during hot periods, and the constraints they have to deal with. In summer, occupants can adopt various behavioural strategies to cope with high temperatures. They may operate windows, adjustable solar shading, fans or air conditioning. They may also adjust their clothing or activities, reduce the use of heat-generating appliances, or move within or outside the dwelling [36,39,40]. However, their behaviour may be constrained by internal or external factors, particularly in urban contexts [40,41], where thermal comfort may come into conflict with other needs. Therefore, a specific model is needed for the summer, when occupant behaviour has a very significant impact on the indoor environment. This study aims to address the gaps identified in the literature: modelling OB during the summer period, taking into account the constraints that urban households may face.

This study aims to understand and model the adaptive behaviour of occupants in urban dwellings during summer, using an agent-based approach, in order to assess the contribution of such modelling to the evaluation of building energy performance and the impact of OB representation on summer comfort. This is a first step toward incorporating occupant behaviour into BPS. The model presented therefore proposes a methodology for modelling adaptive actions within the context of a single case study, which is intended to be further developed with future studies. We begin by presenting the development of an ABM, based on eight semi-structured interviews with households, along with the coupling procedure with a TRNSYS simulation environment for the thermo-aeraulic modelling of dwellings. We then illustrate the relevance of representing the dynamics of occupant adaptive behaviour in BPS through a case study of a retrofitted dwelling.

2. Materials and Methods

To understand the adaptive behaviour of occupants in urban dwellings during summer, a qualitative approach was adopted. Semi-structured interviews were conducted with eight households living in apartments located in the city of Lyon, in south-eastern France, during the summer of 2023. Even if a sample size of eight is small for a statistical analysis, this number was deliberately chosen to allow an in-depth understanding. Conducting the interviews directly inside the dwellings, combined with observations made during the visits, made it possible to deeply explore the specific context, the physical constraints, and the real motivations behind the occupants’ actions.

The households were generally composed of young adults, with at least one working member in each. The eight dwellings are located in different multi-family buildings in urban settings. All of them are equipped with natural cross-ventilation through windows, external solar shading devices, and fans. In addition, two dwellings are equipped with fixed air conditioning systems.

The analysis of these interviews highlights the adaptive strategies implemented during summer, as well as the constraints limiting occupant actions. These findings serve as the basis for the conceptual framework of the agent-based model and the definition of behavioural rules for adaptive actions. Further details on the methodology and results of the semi-structured interviews can be found in Hostein et al. [42].

2.1. Conceptualization of the Behaviour Model

The analysis of semi-structured interviews revealed how occupants behave to cope with the heat in their homes during summer [42]. In total, ten heat adaptation measures were identified among which five main strategies consistently used by the households surveyed were selected to build the model. Eleven constraints that limit the implementation of these measures were also identified. This section briefly describes the relevant adaptation behaviours and constraints that have been translated into occupant behaviour (OB) rules for integration into a BPS environment.

Almost all households reported using cross-ventilation through windows for night-time cooling and during the morning and evening, when outdoor temperatures are lower than indoors. Notably, five out of eight households are equipped with thermometers to monitor indoor air temperature, and occupants operate windows based on indoor readings and outdoor temperature forecasts. External solar shading is closed during the day depending on solar exposure on each façade. Fans are used to increase air movement when occupants feel hot during the day, and sometimes at night. Two households occasionally use air conditioning when they feel very hot. In seven out of eight households, occupants reported adjusting their clothing to cope with the heat, typically by wearing lighter and less covering garments indoors during summer, distinct from the outfits worn outside the home. These results are consistent with heat adaptive actions documented in the literature [36,40], including those reported by Hellwig et al. [39], which reinforces the relevance of the behaviours observed in this urban context.

The constraints limiting these adaptive behaviours are primarily linked to the urban outdoor environment. In addition to weather conditions (e.g., wind, rain), factors such as outdoor noise, air pollution, mosquitoes, natural light, insecurity issues and lack of privacy (e.g., direct views from neighbouring buildings) can restrict the use of windows and/or solar shading. Two additional constraints were identified inside the home: equipment noise and other occupants.

To realistically model these behaviours for integration into BPS tools, it is necessary to translate the qualitative interview data into behavioural rules that can be implemented in a numerical model. The resulting model, called MOBAPY (Modelling Occupant Behaviour using an Agent-based approach with PYthon) developed in Python 3.10.2, is presented in the following section.

2.1.1. Conceptual Framework of the MOBAPY ABM

Among existing theories to explain and model occupant behaviour, we adopted the DNAS framework proposed by Hong et al. [43] in 2015, which aims to standardize occupant behaviour models in relation to building energy use. This framework is organized around four key concepts: Drivers—Needs—Actions—Systems. Occupants react to influencing factors (Drivers) to meet their needs (Needs) by taking specific actions (Actions) involving building equipment (Systems). Figure 1 illustrates how the DNAS framework is implemented in the MOBAPY model.

Drivers are the factors that influence occupant behaviour. In this model, they are grouped into three categories: physical variables, fixed context and time. Physical variables refer to environmental inputs that change during the simulation. These include indoor environmental variables provided by the thermo-aeraulic model and outdoor conditions obtained from the weather file. Fixed context refers to information defined prior to the simulation, and kept constant throughout. It includes contextual data related to the household, the dwelling and the urban environment. Each agent represents an occupant belonging to a household, living in a specific dwelling, and located within a defined urban setting. Time evolves at each simulation step, and includes the hour of the day and the day of the week. The month of the year is not considered, as this model focuses exclusively on summer conditions.

Needs are the expectations that occupants seek to satisfy. In this model, they are divided into two categories: constraints and thermal comfort. In the original DNAS framework proposed by Hong et al. [43], needs encompass physical, biological and comfort-related expectations, as well as non-physical needs. In MOBAPY, non-physical, physiological, visual, acoustic and indoor air quality needs are integrated in a simplified form through the notion of constraints. These constraints represent restrictions imposed by the environmental context, equipment availability, or the presence of other agents within the dwelling. They also include personal constraints, such as the occupant’s activity schedule. The thermal comfort component aims to translate personal and physical environmental variables into a thermal sensation. The latter is assessed using the Physiological Equivalent Temperature (PET) index [44].

The choice of PET is supported by the results of a field measurement campaign performed by Hostein [45], which showed good agreement between perceived thermal sensation and PET predictions. It should be noted that this study involved a limited number of participants. It is a preliminary study that does not demonstrate the superiority of PET over other thermal comfort indicators. The adoption of the PET index in MOBAPY should currently be regarded as an experimental approach.

PET is based on the MEMI (Munich Energy-balance Model for Individuals). PET is equivalent to the air temperature at which, in a typical indoor environment (light wind of 0.1 m/s and mean radiant temperature equal to air temperature), the heat balance of the human body is maintained with core and skin temperatures equal to those under the conditions being assessed [46]. The model is based on a two-node model of the human thermoregulation model. The original code [44,47] was recently revised and made available in its entirety [46]. PET is primarily used to evaluate outdoor conditions [48], but has also been applied indoors, as it is valid for both types of environments [49,50,51,52,53,54]. These references are limited and are discussed in Section 4 along with the limitations of PET for this study.

To evaluate thermal sensation, a scale linking PET values to the standard ASHRAE categories must be used. The thresholds used for MOBAPY are derived from a study based on outdoor responses collected in a Mediterranean climate [55]. This scale was compared to a more traditional one based on European responses [56] in our previous field measurement campaign [45].

Actions are the means by which occupants interact with their environment to satisfy their needs. In Hong et al. [43], an action can take the form of an interaction with a system, a movement, an inaction or a complaint. In MOBAPY, actions are categorized into three types: inaction, reaction and anticipation. At each time step, the occupant decides whether to act, based on their thermal comfort. If discomfort is detected, the occupant reacts by interacting with the available systems. Only one action is taken at a time, but another action may be performed the next time step. Inaction occurs when the occupant is already in a comfortable state, or cannot act (e.g., if they are absent, asleep or if all available adaptation solutions are already in place). Finally, anticipation is implemented in the model only in specific situations: when the occupant wakes up, goes to bed, or moves to or from a room. This is based on observations showing that some actions are more frequently performed when entering or leaving a room than when remaining in it (opening/closing windows, turning on/off lights) [43].

Systems are the equipment, mechanisms or measures with which an occupant may interact to restore comfort or satisfaction with their environment [43]. In this model, all heat adaptation solutions are included under the notion of systems, which are divided into two categories: dwelling equipment, the use of which impacts the indoor environment (e.g., windows, solar shading, and air conditioning); and personal solutions that have no direct impact on other occupants (e.g., fans and clothing adjustments). The new state of the former is transmitted to the thermo-aeraulic model, while the latter is only considered in the thermal comfort calculation at the following time step.

2.1.2. Rules of Conduct

As all surveyed households present homogeneous sociodemographic background and reported similar behaviours, only one occupant profile is considered in this first version of MOBAPY. The rules of conduct implemented in MOBAPY are detailed in Appendix A. They are illustrated here through a commented diagram: Figure 2 presents the rules governing an occupant’s reaction to hot discomfort.

The occupant reacts differently to hot and cold thermal discomfort. If the agent is too hot (PET > 26 °C), they first consider adjusting the state of the window(s) in the room they are currently in. The process begins by checking whether a window is present. In this case, if the window is open and it is hotter outside than inside, the occupant closes it. If the window is closed, they check for any constraints that may prevent opening it: the outdoor temperature must be lower than the indoor temperature, and there must be no noise, mosquitoes, pollution or odours. If opening the window is not possible, the occupant then considers modifying the state of solar shading using a similar logic, taking into account direct solar radiation. If the above actions are not available, or have already been implemented and the discomfort persists in the next time step, they then consider adjusting their clothing (in 0.1 clo increments) up to a certain value (underwear or tank top depending on constraints), followed by activating the fan, and, as a last resort, using air conditioning. A hierarchy is applied to manage potential conflicts when multiple agents (i.e., occupants) disagree on actions at a given time step.

Fans are modelled as mobile agents with three power levels. They only influence perceived comfort in MOBAPY, and not sensible heat gains and electricity use. The resulting air speed is based on the Givoni diagram zones: 0, 0.5, 1 and 1.5 m/s. Fans can be moved between rooms when requested by an occupant. The default behaviour is to start with the lowest power setting and progressively increase the level at subsequent time steps if discomfort persists. When the number of fans is limited, the occupant who turned on the fan retains exclusive control over it until it is turned off. However, other occupants may benefit from the elevated air speed if they are in the same room.

Finally, agents only activate the air conditioning if their PET exceeds 28 °C, reflecting the behaviour reported by interviewed occupants in dwellings equipped with AC, who stated they only used it (set at 27 °C) when they felt sufficiently hot. AC is turned off as soon as PET falls below 26 °C, corresponding to a neutral thermal sensation. Other rules of conduct—such as responses to cold or anticipatory actions based on the time of day—are detailed in Appendix A.

2.2. Integration of MOBAPY into a Coupled Simulation Approach

2.2.1. Technical Implementation of MOBAPY in Python

MOBAPY has been developed in Python for several reasons: the widespread use and popularity of the language, the availability of numerous libraries, online support, its object-oriented programming capabilities, and the existence of several TRNSYS types that allow for coupling with Python scripts (and more generally its ease of integration with various BPS tools).

Several agent-based libraries are available in Python, including Mesa [57,58], AgentPy [59], and Melodie [60]. AgentPy (ap) was chosen for this work due to its Python-native development, ease of use, ability to simulate network-based environments, and well-documented framework with example code [61]. MOBAPY was thus developed using an object-oriented programming approach. Its UML class diagram is provided in Appendix A. The System and Occupant classes inherit from the Agent class in the AgentPy library. The ResidentialModel, Dwelling and Room classes inherit respectively from the Model, Network and AgentNode classes of AgentPy. These class attributes are used to update environmental variables based on the outputs of the dwelling’s thermo-aeraulic model, and to define the behavioural rules governing occupant conduct.

At each simulation time step, the Python code executes four processes corresponding to the DNAS framework: updating the Drivers, calculating the Needs of each occupant, determining Actions, and updating the state of Systems. In the first step, occupants are assigned to rooms according to the occupancy schedule and the environmental variables for each room are updated. The second step consists of checking the relevant constraints and calculating the thermal sensation using the PET index, computed with the pythermalcomfort package [62]. If, in step 3, occupants decide to take action, the corresponding system states are updated in step 4.

2.2.2. Constraints

As indicated above, occupants may be subject to constraints that compromise their ability to maintain thermal comfort (Figure 1). In MOBAPY, ten types of constraints are implemented (

Table 2. Attributes representing constraints in MOBAPY.

Constraint	Python Class	Used Name	Nature	Variable or Parameter Description
Pollution	Room (Outdoor object)	pollution	Planning	Outdoor environment polluted
Mosquito	Room (Outdoor object)	mosquito	Planning	Presence of mosquito in the external environment
Noise	Room (Outdoor object)	noise	Planning	Outdoor environment noisy
Fan	ResidentialModel	number_fans	Fixed	Number of fans in the dwelling
Air conditioning	Room	availability	Fixed	Existence of air conditioning in the room
Direct views	Room	vis-a-vis	Fixed	Presence of direct views for the room
Insecurity	Room	insecurity	Fixed	Feeling of insecurity in the room
View	Occupant	want_view	Fixed	The occupant wants to see outside
Privacy	Occupant	want_privacy	Fixed	The occupant does not want to be seen from outside
Darkness	Occupant	want_darkness	Fixed	The occupant wants to sleep with the shutters closed

). These constraints are binary in nature. Prior to each simulation, the constraints to which agents are exposed must be specified. A given configuration corresponds to a specific combination of these constraint settings. Each agent represents an occupant living in a dwelling situated within a specific environment. Accordingly, constraints may relate to the occupant (e.g., need for privacy), the dwelling (e.g., equipment availability), or the external environment (e.g., noise, pollution).

We consider pollution caused by road traffic emissions during rush hour. During these times, occupants cannot open their windows from 7 a.m. to 10 a.m. and from 4 p.m. to 7 p.m. on weekdays [63]. If the occupants consider that there are mosquitoes in the surrounding area, they cannot open the windows from 5 p.m. to 10 p.m. every day, which corresponds to the peak activity period of the tiger mosquito [64]. We consider a constant noise nuisance, which therefore prevents night-time cooling. The other constraints are also constant. Occupants may be visible to neighbours inside their home (vis-a-vis). In that case, they will not be able to wear extremely light clothing if the solar protections are open. If they feel unsafe (insecurity), occupants will not be able to leave both a window and its solar protections open when leaving a room.

2.2.3. Interoperability with BPS

To simulate the dynamic thermo-aeraulic behaviour of a dwelling and the thermal comfort of its occupants during summer, while accounting for both the immediate external environment under different climatic scenarios and the interactions with occupants, MOBAPY must be coupled with other tools. The TRNSYS simulation environment was chosen, in combination with CONTAM, to model the thermal and airflow dynamics. Figure 3 presents the interoperability diagram between the different components (tools, input and output files) involved in this coupled approach.

This figure illustrates the feedback loop between the occupants and their environment, capturing the dynamic nature of behaviour over time. A time step of 10 min is used as a compromise between computational cost and the resolution required to capture the dynamics of thermal and aeraulic variables. Each simulation is configured with input information on occupant constraints and a meteorological file.

Each component is modelled by a ‘Type’ from the TRNSYS library. Type 15 reads weather data from EPW-format files. Type 56 simulates the hygrothermal behaviour of a multi-zone building. Mean radiant temperature used in PET is the weighted mean surface temperature of each zone, a method shown to exhibit lower errors than the zone-averaged approach, particularly near glazed surfaces [65]. It is coupled with Type 98, which computes the aeraulic behaviour of the multi-zone building using CONTAM. Type 3157 enables the coupling with the MOBAPY occupant behaviour model developed in Python.

2.3. Application of the Developed Toolchain: Description of the Case Study

To demonstrate the value of representing OB in building simulation, the proposed approach is applied to a case study of a retrofitted urban multi-family dwelling previously presented in [66]. It is a 65 m² two-bedroom apartment located on an intermediate floor. It has two façades facing east and west, enabling natural cross-ventilation. It also has external solar shading. The building features high external thermal insulation and strong thermal inertia. A thermo-aeraulic model of the dwelling was developed, using TRNSYS and CONTAM, calibrated and validated against experimental field measurements. A set of 25 meteorological files was generated to represent the future climate of an urban district during summer and heatwave periods. As described in [66], these future weather files are derived from regional climate projections and the methodology generates complete meteorological files in which all weather variables (air temperature, humidity, wind speed, solar radiation, etc.) remain physically consistent with one another. This is essential for reliable BPS. Moreover, the method used ensures that the generated files include the intensity, duration, and temporal evolution of projected heatwaves, which are the critical periods for assessing overheating risk. The historical scenario is based on a 30-year reference period (1976–2005), consistent with standard climate modelling conventions (EURO-CORDEX). The approach also incorporates the characteristics of the urban microclimate.

Further details on the dwelling geometry, thermal characteristics and thermo-aeraulic model validation are also provided in [66].

In this study, the dwelling is occupied by a hypothetical household consisting of two adults working during the week, one of whom works from home in the office room on Tuesdays and Thursdays. Activity level depends on the activity of each occupant [67] considered constant for each room (0.8 met bedroom, 1 met living-room, 1.2 met office, 1.6 met bathroom and kitchen). The interior doors are assumed to be always open. MOBAPY is used to model the state of the windows, solar shading devices, air conditioners, fans and clothing adjustments. If a window is open while its solar shading device is closed, the effective opening area is reduced by half. To implement a household model in MOBAPY, it is necessary to enter the number of rooms and the equipment they contain, the number of occupants, and specify whether or not there are constraints at the level of the dwelling, each room, and each occupant (Table 2).

Once the models are coupled as described in Section 2.2.3, the evolution of the indoor environment and occupant behaviour is simulated under different constraint configurations, as defined in Section 2.2.2.

Indoor environment conditions are assessed from 1 June to 30 September (simulation starting two weeks earlier), using thermal comfort indicators based on the percentage of time (PT_PET) and degree-hours (DH_PET) during which the PET exceeds 28 °C, for a total of 2241 occupied hours. The risk indicator used in [66] is also applied here: the maximum number of successive days during which the bedroom temperature fails to drop below 26 °C (iT_26°C), representing the occupants’ reduced capacity to recover from prolonged heat exposure.

3. Results

3.1. Dynamics of Adaptive Behaviour in an Unrestricted Configuration

Figure 4 presents the simulation results from the coupled models for the living room over one week, from 25 June to 2 July, under an unrestricted configuration, where all constraints listed in Table 2 are set to false. The figure displays the evolution of indoor and outdoor air temperatures, alongside occupants’ presence and their actions on the windows, solar shading, fan and AC. The meteorological file used for this example is the one constructed from IPSL/RCA4 model data between 2070 and 2099 with RCP 8.5 [66].

During this week, the outdoor air temperature fluctuates between 15 °C and 40 °C, with very hot conditions during the first three days when night-time temperatures do not fall below 20 °C. Occupants are globally present in the living room during the evenings, and throughout the day on the weekend.

Indoor air temperatures range from 18 °C to 32 °C, with higher values during the hottest days. The figure illustrates the impact of adaptive actions on indoor temperatures, particularly window opening. Windows are opened at night, at the same time as solar shading devices. The latter are opened during the day when at least one occupant is present in the room and there is no longer any direct solar radiation. In this case, the windows can also be opened if the indoor temperature exceeds the outdoor temperature. Windows and solar shading devices are typically closed each morning when the occupants wake up. Although they may not be present in the living room, they are present elsewhere in the dwelling. The fan is used when occupants feel hot, but only when they are present in the room. Air conditioning is used as a last resort and is therefore rarely turned on. Similarly, it is only turned on when occupants are present.

3.2. Impact of Occupant Behaviour Under Various Constraint Configurations

To assess the impact of behavioural adaptation on summer comfort for urban dwelling occupants, several constraint configurations were simulated with MOBAPY. Five climate scenarios were considered, including a historical climate and future climates for the periods 2035–2064 under 2070–2099 and RCP 4.5 and 8.5 [66]. The RCP 4.5 and RCP 8.5 scenarios are the most commonly used in climate change impact studies [68]. The RCP 4.5 scenario is considered the most probable [69]. The RCP 8.5 scenario is the most pessimistic and represents the future climate if nothing changes [70].

Figure 5 illustrates the thermal comfort results for a non-air-conditioned dwelling, focusing on the two constraints that have the most significant impact: absence of a fan, constant external noise, as well as their cumulative effect.

Each graph corresponds to the bar chart of a thermal comfort indicator. The configurations are indicated on the x-axis, while the five climate scenarios are distinguished by different colours. The height of each bar represents the average value of the indicator across five combinations of climate models, and the black error bars indicate the range (the results obtained with the hottest and coldest climate models).

It was chosen to focus on the two most critical constraints identified (the absence of a fan and constant outside noise) since they affect different adaptive actions, and thus, their combination further exacerbates thermal discomfort. Even under historical climate conditions, the PT_PET indicator shows an average of 21% of uncomfortable occupied time for dwellings without fans and exposed to constant noise. This percentage increases to 68% under the worst-case climate scenario (2070–2099 RCP 8.5). In contrast, in an unrestricted configuration, the average PT_PET remains around 20% under the same climate scenario.

DH_PET exhibits a similar trend, with more pronounced results under constrained configurations. The iT_26°C risk indicator highlights the benefits of night-time window opening, which helps to reduce the duration heat peaks. Without the noise constraint, the average iT_26°C remains below 6 days, except for the most extreme climate scenario (14 days on average). However, when night-time ventilation is restricted due to noise, the indicator value doubles. Notably, under the worst climate scenario, there is substantial variability between the climate models.

These results underline the importance of accounting for constraints affecting occupant behaviour when assessing the summer comfort performance of dwellings.

4. Discussion

The model presented in this article is an initial version based on a limited number of semi-structured interviews. It is intended to be expanded upon through further research, particularly on the following points. A comparison with in situ measurements is one of the approaches missing from this study, which could be one of the key priorities for future work.

MOBAPY accounts for occupant adaptive behaviour regarding five systems: windows, solar shading devices, clothing, fans and air conditioning. However, other adaptive actions reported during the interviews, such as taking showers, limiting the use of heat-generating appliances, or temporarily leaving or isolating parts of the dwelling, were not integrated into the current version of the model [42]. To limit computational complexity, a fixed occupancy schedule was also used. A potential improvement would be to introduce flexibility into this schedule, enabling the representation of a wider array of adaptive strategies. Additionally, certain environmental constraints, such as wind or rain, were not considered, though they could be incorporated using inputs from meteorological files.

This approach was applied to a household of two occupants living in a two-bedroom apartment. Thanks to its object-oriented programming architecture, MOBAPY can accommodate a variable number of rooms and occupants without requiring code modification. This structure also facilitates future extensions.

Thermal sensation in MOBAPY is evaluated using the Physiological Equivalent Temperature (PET) index, based on the MEMI thermo-physiological model, with thresholds derived from a study based on outdoor responses. However, the model could be adapted to use any other thermal comfort indicator that incorporates the six fundamental parameters of comfort. The methodology adopted here aligns with a hybrid comfort modelling approach [71], i.e., based on adaptive comfort theory while explicitly representing adaptation processes. Physiological adaptation is considered thanks to the MEMI thermo-physiological model embedded in the PET calculation, and behavioural adaptation is integrated through the dynamic coupling with MOBAPY. However, psychological adaptation is not yet addressed. Incorporating this aspect, as done in the ATHB model [72], would represent a valuable extension.

It should be noted that the PET index currently implemented in MOBAPY is primarily used for outdoor environments. PET has also been applied for indoor studies [49,50,51,52,53,54] although their objectives differed from this study’s (transition between thermal comfort and discomfort). Two-thirds are focused on heat stress issues rather than thermal comfort [49,50,52,53]. Two doctoral dissertations concern the French context [49,52], whereas the other studies examine different types of climates. Two studies cover educational buildings [53,54], not residential buildings as this work does. Two studies focus on indoor and outdoor conditions [49,54], and Jiang et al. [51] explores short-term dynamic thermal perception under transient indoor–outdoor conditions.

Further work is needed to demonstrate that PET accurately represents occupants’ thermal perception in indoor environments. The choice of this indicator was based on a preliminary study with a limited number of respondents. New experimental campaigns involving different types of households, buildings, and climates are thus needed to validate it. Future research plans to explore the use of other comfort indicators in MOBAPY (such as PMV, aPMV, or ePMV, SET, ATHB, or DTS) as well as the associated thresholds that trigger occupant actions. These indicators consider the effects of temperatures, humidity, air velocity, clothing, and metabolism on thermal comfort. The PMV is one of the most widely recognized and commonly used indicators in the field of building science, and is included in the ISO 7730 standard for air-conditioned buildings [67]. Since some studies have found that it tends to overestimate hot discomfort [73], an adaptative approach is preconized for naturally ventilated building in EN 16798-1 [74] and ASHRAE 55 [75] standards, by establishing a link between the moving average outdoor temperature and the indoor operative temperature range considered comfortable. However, this simplified approach considers only the indoor operative temperature. Indicators applying statistical corrections on PMV have been developed to implicitly incorporate adaptive mechanisms, such as aPMV [76] or ePMV [77]. SET is used more frequently in indoor environments than PET and is based on the Gagge model [78]. The latter has recently been improved to calculate a Dynamic Thermal Sensation (DTS) [79]. Finally, ATHB combines the adaptive approach with the heat balance method [72].

In its current form, MOBAPY simplifies the consideration of non-thermal comfort dimensions (e.g., physiological, visual, acoustic, and indoor air quality) by integrating them merely as binary constraints. A next step would be to develop specific models for these comfort domains, similar to the thermal comfort model, to enable a more comprehensive and subjective representation of occupant perception. Similarly, while factors like noise and air pollution in the urban environment are currently treated as static daily scenarios, they could eventually be modelled as dynamic, continuous variables.

The adaptive actions themselves are also accounted for using a simplified approach (windows and solar shading devices open/closed, clothing increments of 0.1 clo, fan speeds corresponding to Givoni zones, air conditioning on/off). One potential future development for MOBAPY could be working on refining these actions with percentage of opening, clothing sets, setpoint temperature variations, etc. Figure 4 shows the changes over time—spanning one week—in temperature and some adaptive actions illustrated in simplified form, for a single model configuration without constraints. The goal is to provide insight into how the current model works. It might be interesting to examine the details (changes in fan speed, clothing increments, and PET for each occupant) or other configurations to improve the model. Data is available from the existing simulation outputs.

Only a single occupant profile is currently implemented in MOBAPY, based on a unified set of behavioural rules. This simplification is justified by the homogeneous sociodemographic background and declared behaviours of the eight interviewed households, all of whom reported environmentally conscious attitudes, particularly regarding air conditioning use. Extending the model to include a broader diversity of profiles would be beneficial. It could be achieved by conducting interviews with a larger and more diverse sample of households. It would also be relevant to apply the methodology to other building types and urban contexts.

Although one profile was implemented based on eight semi-structured interviews, a large-scale socioeconomic survey (1072 respondents) was conducted as part of the CREATIV project [80]. It allows for the heterogeneity of housing and households to be contextualized within a representative sample. Overall, households adjust their windows more frequently than their shutters, followed by fans, and lastly air conditioning (no questions were asked about clothing). This observed frequency hierarchy perfectly mirrors the sequence of actions implemented in MOBAPY’s decision tree. The analysis also identified three main clusters. The second cluster (31.2% of the sample) consists of households that adopt a ventilation-based adaptation strategy, whose behaviour is similar to that reported in the eight semi-structured interviews in our study. This large-scale survey thus supports the actions implemented in MOBAPY, which are based on reports from eight households belonging to the second cluster. Future work could include the integration of the two other clusters.

MOBAPY is only applied to one case study: a highly insulated, cross-ventilated apartment occupied by a hypothetical two-adult household with a fixed schedule. Results are expected to differ for other building typologies. A less insulated building would likely be more affected by solar protections use. The effectiveness of night-time ventilation will also be significantly reduced in single-sided dwellings or those equipped with window screens (which restrict airflow). Finally, this household profile is particularly active in terms of summer thermal comfort; the impact of OB is expected to be less significant for other profiles.

5. Conclusions

Developing behavioural models for summer or hot weather is a major research area for assessing the efficiency of occupant-dependent passive cooling strategies. Agent-based systems are particularly well suited to represent the complexity of occupants’ overall behaviour and to model their adaptive actions, thereby allowing for the assessment of their impact on buildings’ thermal performance. Several agent-based models for the building context have been developed in other studies, focusing mainly on energy consumption, and are therefore less adapted to study thermal management strategies employed by occupants during hot periods. To define rules of conduct adapted to this context, semi-structured interviews were conducted with households in Lyon during the summer of 2023. Based on these results, an agent-based model called MOBAPY was developed to represent the summer behaviour of occupants in an urban dwelling. It models the use of windows, solar shading devices, air conditioning, fans and clothing under different constraints (external disturbances, configuration of the dwelling and availability of equipment, personal preferences). It is coded in Python, a widely used programming language, and can be easily coupled with building performance simulation tools.

MOBAPY is coupled with a thermo-aeraulic dwelling model built with TRNSYS and CONTAM to illustrate the value of representing adaptive behaviour. The evolution of the indoor environment is simulated under several climatic hypotheses with different configurations, representing constraints impacting the adaptive behaviour of the occupants. The results show that occupant behaviour is crucial in energy-efficient dwellings. For the worst climate scenario, PT_PET can fall to 20% without behavioural constraints, whereas it reaches 68% with constraints. This contrast raises a critical question: is it truly possible for occupants to act in the most thermally efficient manner, taking into account personal or external constraints, particularly in an urban context? Simulations under different constraint configurations reveal the beneficial effects of opening windows: the confinement caused by constant external nuisance leads to long periods during which the indoor temperature remains high (for this well-insulated case study dwelling). A fan is also a solution that greatly reduces thermal discomfort, thanks to the increased air speeds it generates.

Beyond the specific findings of this case study, the application shows that the representation of occupancy behaviour has a strong impact on the evaluation of the building’s thermal performance and possible energy consumption linked to the use of air conditioning. Carefully representing occupant behaviour, according to their profile and the constraints to which they are subject, is therefore a major challenge in assessing the resilience of buildings to climate change and establishing robust adaptation strategies.