Abstract
Residential buildings account for a substantial portion of global energy consumption and are critical for ensuring safe and comfortable indoor environments. Accurate prediction of indoor temperature and humidity is essential for thermal comfort, energy efficiency, and occupant health, yet remains methodologically complex. This review synthesizes 91 peer-reviewed studies published between 2019 and 2025 that applied white box (physics-based) and black box (data-driven) modeling approaches to residential buildings. Studies were identified through Google Scholar and Web of Science, screened using defined inclusion and exclusion criteria, and evaluated based on model type, predictive variables, validation method, and performance metrics. White box models, including nodal, zonal, computational fluid dynamics (CFD), and hybrid physics-based frameworks, capture mechanistic heat and moisture transfer processes and have evolved toward more integrated hygrothermal and airflow coupling since 2021. Black box methods, including shallow and deep neural networks, regression models, and hybrid or ensemble architectures, demonstrate high predictive accuracy for short-term and real-time applications, often achieving sub-degree temperature errors and increasingly predicting both temperature and humidity. Comparative findings show that temperature remains the predominant predictive variable, while humidity, though vital for comfort and health, is less frequently modeled. Geographically, studies are concentrated in Europe, East Asia, and North America, with limited representation of tropical and Global South regions. Seasonally, most research has emphasized heating conditions, though recent efforts address cooling, overheating, and mixed-mode ventilation. Evaluation remains fragmented across metrics such as RMSE, MAE, and correlation coefficients, underscoring the need for standardized reporting practices. Overall, physics-based and data-driven approaches are complementary. The former provide interpretability and physical realism, while the latter offer adaptability and scalability. Future research should prioritize coupled heat–moisture modeling, extend analysis to underrepresented climates and housing types, and develop unified validation benchmarks. Advancing along these directions will strengthen predictive reliability and support healthier, more energy-resilient residential environments.
Keywords: Energyplus, Computational fluid dynamics, Physics based modeling, Data driven modeling, Machine learning, Indoor humidity, Indoor temperature, Residential Buildings
1. Introduction
The Intergovernmental Panel on Climate Change (IPCC) reports that human-induced climate change has led to an increase in the frequency and intensity of extreme thermal events across most land areas since the 1950s. Projections indicate that with each additional 0.5 ° C of global warming, the occurrence and severity of heat extremes will continue to rise. For instance, a heatwave that was considered a once-in-a-decade event in pre-industrial times is now occurring approximately 2.8 times more frequently. If global temperatures rise by 2 ° C, such events are projected to occur 5.6 times per decade. These findings underscore the escalating impact of climate change on extreme thermal events worldwide [1]. Wu and Zhou [2] estimate a ten-fold increase in heat wave deaths in the Eastern U.S. by 2057–2059, with 1400 to 3600 deaths annually by 2058. Numerous studies have confirmed the strong link between extreme heat and adverse health outcomes, particularly among older adults who spend most of their time indoors and may lack air conditioning [3–6]. These findings underscore the importance of considering and understanding indoor overheating, which the Chartered Institution of Building Services Engineers (CIBSE) defines using three criteria: hours of exceedance, daily weighted exceedance, and upper limit temperature. A building is considered at unacceptable risk of overheating if it fails any of these criteria [7,8]. This overheating results in thermal discomfort, reduced productivity, and health risks, particularly for vulnerable populations like older adults and low-income households who may struggle to adapt [4,9–11].
Despite reports of high energy used to cool and heat up residential buildings, achieving thermal comfort is often elusive. For instance, Pathan et al. [12] reported a significant risk of overheating in 37 % of living rooms and 49 % of bedrooms during summer months, highlighting the need for adaptation strategies in residential building regulations, design, and retrofit.
To develop effective adaptation strategies, accurate prediction of indoor temperature and humidity levels is crucial. Indoor thermal conditions are influenced by numerous factors, including building materials, occupancy behavior, ventilation strategies, and external climatic conditions. The building envelope’s insulation and airtightness, ventilation rates, and thermostat settings significantly impact energy consumption and thermal comfort [13]. Occupant activities, such as movement, opening and closing of doors and windows, and use of appliances, also affect the indoor microclimate [14]. Without robust modeling, building designers and policymakers lack the tools to anticipate how buildings will respond to climate extremes, leading to inefficient adaptation efforts and increased vulnerability to heat-related health risks. Predictive modeling enables the assessment of cooling system strategies and performance, ensuring that interventions are both effective and energy efficient.
While several existing studies have evaluated modeling approaches for building energy and indoor environment prediction, most have focused on general building sectors rather than residential dwellings. For instance, Foucquier et al. [15] and Zhao et al. [16] provided broad comparisons of white box, black box, and hybrid models, highlighting methodological strengths and limitations but without emphasis on housing. Yu et al. [17] focused on uncertainties in prediction models, while Chen et al. [18] underlined the advantages of hybrid methods when data or physical information are limited. Reviews such as Hamdaoui et al. [19] addressed hygrothermal modeling of building materials, and Al Mindeel et al. [20] synthesized multi-objective optimization studies spanning energy, comfort, and indoor air quality. Others concentrated on specific approaches: Lu et al. [21] examined ANN models, Ngarambe et al. [22] and Feng et al. [23] explored ML for thermal comfort prediction, and Saini et al. [24] reviewed AI methods for indoor air quality. Broader works, including Swan et al. [25], assessed end-use energy in the residential sector, but with emphasis on macro-level consumption rather than micro-level prediction of indoor conditions.
Collectively, earlier reviews have advanced our understanding of building energy and environmental modeling, yet residential dwellings remain substantially underrepresented. While commercial and institutional buildings are typically characterized by stable occupancy patterns, standardized HVAC operation, and extensive monitoring infrastructure, homes exhibit far greater variability in occupant behavior, construction era, and environmental control. These factors introduce substantial uncertainty into the prediction of indoor temperature and humidity, underscoring the need for a targeted synthesis of residential-focused modeling approaches.
This review distinguishes itself from prior work by concentrating exclusively on residential buildings and by emphasizing the combined prediction of indoor temperature and humidity, a critical but often overlooked aspect of indoor environment research. It systematically compares physics-based (white-box) and data-driven (black-box) modeling techniques, examining how each captures coupled thermal and moisture dynamics within homes. Unlike broader reviews spanning multiple building types or timeframes, this study focuses on the 2019–2025 publication window to capture recent methodological advancements, including coupled heat–moisture transfer simulations, hybrid physics-informed frameworks, and AI-enhanced calibration and validation strategies. Through this focused scope, the review contributes an updated synthesis of how contemporary modeling methods represent indoor environmental behavior in residential contexts. It further highlights the implications of these modeling practices for thermal comfort, occupant health, and energy resilience, providing a distinct and timely advancement beyond previous generalized assessments of building energy and environmental modeling.
To guide this review, we focus on the following research questions:
What white box and black box modeling approaches have been applied to predict indoor temperature and humidity in residential buildings?
How are these studies distributed geographically, seasonally, climatically, and across different residential building typologies over the review period?
What predictive variables and performance metrics have been adopted, and how do they vary across modeling approaches?
What are the key strengths, limitations, and methodological trends of these modeling techniques?
The structure of this paper is as follows: Section 2 outlines the methodology, including search strategy and inclusion/exclusion criteria. Section 3 examines the geographic, seasonal, and climate-zone distribution of the reviewed studies. Sections 4 and 5 provide detailed analyses of white box and black box modeling approaches, respectively, discussing model types, performance, predicted variables, input variables, building typologies, temporal trends, and methodological limitations. Section 6 synthesizes the findings, highlighting the applications of both paradigms, the distribution of predictive variables and performance metrics, and the strengths and weaknesses of each approach, before presenting future research directions.
2. Methodology
This review focuses on exploring indoor temperature and humidity modeling strategies in residential buildings using white box and black box approaches. The review aims to evaluate how these modeling strategies address thermal comfort and predictive accuracy under various climatic and occupancy conditions. Special emphasis is placed on the strengths, limitations, and practical applicability of these models in residential contexts.
Thermal comfort, a primary consideration in indoor environmental quality, is defined by American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) as “that condition of mind which expresses satisfaction with the thermal environment” [26]. This state is influenced by six primary factors: air temperature, radiant temperature, humidity, air speed, metabolic rate, and clothing insulation [26]. Given the evolving understanding of personalized thermal comfort, recent models have adopted personal comfort approaches, utilizing wearable technology and environmental sensors to account for individual preferences [27]. This review contextualizes these advancements by examining the role of accurate indoor temperature and humidity predictions in delivering optimized comfort solutions for diverse residential typologies.
2.1. Search strategy
A comprehensive search was conducted using Google Scholar to identify relevant studies. The Boolean string used was:
(“indoor temperature” OR “indoor humidity”) AND house AND (“white box model” OR “physics-based model” OR “physical modeling” OR ”building energy simulation” OR “black box model” OR “data-driven modeling” OR “machine learning” OR “artificial neural network” OR “support vector machine” OR “random forest”) AND (EnergyPlus OR DesignBuilder OR TRNSYS OR MATLAB OR CFD OR WUFI).
To complement this, Web of Science was queried with the following Topic (TS) search:
TS=((“indoor temperature” OR “indoor humidity”)
AND (“residential building*” OR house* OR “housing unit*” OR “single-family” OR “multi-family”)
AND ((“white box model*” OR “physics-based model*” OR “physical modeling” OR “building energy simulation” OR EnergyPlus OR DesignBuilder OR TRNSYS OR WUFI OR CFD OR MATLAB)
OR (“black box model*” OR “data-driven model*” OR “machine learning” OR “artificial neural network*” OR “support vector machine*” OR “random forest*” OR “regression model*”))
AND (“thermal comfort” OR “humidity control” OR “indoor climate prediction”)).
This structure was designed to capture studies modeling indoor temperature or humidity in residential settings across both white box and black box approaches.
2.2. Inclusion and exclusion criteria
2.2.1. Inclusion criteria
Studies that model or predict indoor temperature and/or indoor humidity (or directly derived thermal-moisture metrics) in residential buildings.
White box / physics-based models (nodal, zonal, CFD, hygrothermal, etc.), including hybrids within physics-based approaches (e.g., zonal–CFD coupling, WUFI–Modelica).
Black box / data-driven models (machine learning, artificial neural network, support vector machine, random forest, regression, etc.), including hybrids within data-driven approaches (ensembles, stacked models).
Publication year: 2019 to 2025.
Peer-reviewed journal articles or full conference papers, published in English.
Papers must provide clear methodological description and quantitative results and/or validation (e.g., error metrics, measured data comparison, sensitivity or uncertainty analysis).
2.2.2. Exclusion criteria
Studies that do not report indoor temperature or humidity outcomes (energy-only, IAQ-only without thermal/moisture metrics).
Buildings that are not residential (commercial, industrial, large public, etc.) unless a mixed or distinct test case is analyzed.
Insufficient methodological detail or absence of quantitative evaluation/validation.
Cross-paradigm hybrid studies that integrate physics-based and data-driven models in a single framework.
Non-English, abstracts only, posters, theses (unless a peer-reviewed version exists).
Pure review papers (use for context only, not counted as primary studies).
As shown in Fig. 1, the search produced 4938 records in total, with 4756 from Google Scholar and 182 from Web of Science. Screening based on the exclusion criteria removed items outside the review scope, including studies on non-residential buildings, those limited to outdoor climate, papers published outside the year range, or those without indoor temperature or humidity outcomes. After this step, 81 relevant papers were retained from Google Scholar. The Web of Science records were screened in the same way and then merged with the Google Scholar set, giving 178 papers for full-text assessment. From this pool, 77 studies were excluded due to insufficient methodological detail, duplicate entries, missing temperature or humidity results, or reliance on methods that did not fit the review categories. This left 91 studies that met all inclusion criteria, consisting of 71 physics-based (white box) and 30 data-driven (black box) studies, all published between 2019 and 2025 with clear methodology and quantitative validation.
Fig. 1.
Flow chart showing papers included and excluded from the review.
3. Geographic, seasonal, and climatic patterns of studies
The country-level distribution of the reviewed studies highlights geographical disparities in research output and data accessibility, providing broader context for regional research capacity and gaps in residential indoor climate modeling. In the worldwide map of the study origins as presented in Fig. 2, East-Asian and North-Atlantic shows a strong presence. China alone supplies one-fifth of all evidence, followed by the United States, Spain, Poland, and Japan. This clustering is more than a bibliometric curiosity. It mirrors where housing stock faces pressing questions about extreme heat, heating electrification, and rapid retrofit programs. For example, Chinese investigations stretch from high-altitude Tibetan dwellings requiring heavy insulation [28] to district-heated high-rise apartments in Tianjin that must balance summer humidity against winter load shifting [29]. In the United States, research pivots toward smart-thermostat analytics and model predictive control, as seen in the multi-state dataset analyzed by Huchuk et al.’s [30] and in Alhamayani et al.’s [31], a midwest case study. Southern European work, led by Espinosa et al. [32] and Vila-Hernández et al. [33], focuses on mixed-dry Mediterranean housing where passive shading and green-roof retrofits dominate design discourse.
Fig. 2.
Global spread of reviewed studies.
When these countries are re-sorted into broad climate bands (Fig. 3), a second pattern emerges. Mixed climates account for roughly half of all studies, a consequence of the dominance of ASHRAE zone 4A (Mixed-Humid) as detailed in Table 1. Here, both summer overheating and winter heating loads matter, which explains the popularity of wholeyear physics-based simulations such as Sarna et al.’s [34] calibrated passive house in Poland or Martin et al.’s [35] survey of occupied homes across the US Mid-Atlantic. Warm and tropical bands contribute a combined quarter of the dataset, often highlighting low-energy cooling measures. Malaysian terrace-house work by Murtyas et al. [36] and Singapore façade optimization by Tong et al. [37] showcase typical concerns, high humidity, natural ventilation, and roof reflectivity, found in zones 0A, 1A, and 2B Cool-humid and cold-humid zones are smaller but distinctly Nordic-Canadian, with Farahani et al. [38] predicting heat-wave impacts on Finnish apartments, while Delcroix et al. [39] couples neural models to manage peak heating in Quebec’s all-electric houses.
Fig. 3.
Distribution of studies by Climate Zones.
Table 1.
Estimated Counts of studies by ASHRAE climate zone (0A–8).
| ASHRAE Zone | Description | Studies |
|---|---|---|
| 0A | Equatorial - Very-Hot / Humid | 6 |
| 0B | Equatorial - Very-Hot / Dry | - |
| 1A | Tropical - Hot / Humid | 2 |
| 1B | Tropical - Hot / Dry | 2 |
| 2A | Tropical / Sub-Tropical - Warm-Humid | 4 |
| 2B | Tropical / Sub-Tropical - Warm-Dry | 6 |
| 3A | Warm-Humid | - |
| 3B | Warm-Dry | 4 |
| 3C | Warm-Marine | 12 |
| 4A | Mixed-Humid | 45 |
| 4B | Mixed-Dry | 1 |
| 4C | Mixed-Marine | - |
| 5A | Cool-Humid | 6 |
| 5B | Cool-Dry | 1 |
| 5C | Cool-Marine | - |
| 6A | Cold-Humid | 6 |
| 6B | Cold-Dry | - |
| 7 | Very-Cold | 2 |
| 8 | Sub-Arctic | - |
Seasonality cuts across these spatial observations. Data-driven papers tend to zoom in on one dominant season. Sözer et al. [40] calibrate a short-term winter model for an Istanbul elderly-care block, reflecting a heating-season design priority in zone 3C. By contrast, Farahani et al.’s [38] summertime forecast tool for Nordic apartments targets overheating episodes that have become more frequent even in cool regions. Physics-based studies still report annual metrics but often separate summer discomfort hours from winter loads, as in Winkler et al.’s [41] humidity-sensitive comfort model and Li et al.’s [42] plateau solar house case. The mixed-humid belt thus acts as a methodological bridge, combining season-specific insights from both hot and cold climates.
Gaps are as instructive as concentrations. Marine sub-zones (3A, 4C, 5C) and very-hot dry 0B appear almost blank in Table 1, yet they include megacities such as Houston, Sydney, and Riyadh. Likewise, sub-arctic zone 8 remains uncharted despite growing construction in northern Canada and Scandinavia. Only two studies venture into very-cold zone 7, Vadiee et al’s [43] Swedish apartment heating analysis and Laukkarinen et al.’s [44] Finnish single-family long-term temperature forecast, suggesting that resilience research in extreme cold is still nascent. Addressing these geographic and climatic blind spots should be a priority for the next wave of indoor climate modelling projects. Further details on these individual studies are summarized in Appendix Table A and B.
4. White box modeling
Whitebox or physics-based models simulate the thermal behavior of different building types and their characteristics by employing fundamental physical principles. Components typically modeled include space heating, natural ventilation, air conditioning systems, passive solar heating, photovoltaic panels, hygrothermal effects, occupant behavior, and climatic conditions [15]. For a detailed overview of the reviewed studies applying these methods, see Appendix Table A.
4.1. Model types: performance, predicted variables, and building types
4.1.1. Nodal model
Nodal models simulate building thermal environments through lumped capacitance resistance networks, where each thermal zone is represented as a well-mixed node of air and surface interactions. This simplification ensures computational efficiency and makes nodal models ideal for long-term simulations such as annual energy analysis, overheating assessments, and parametric evaluations across different residential building types.
As shown in Fig. 4 and detailed further in Appendix A, most nodal studies focused on predicting indoor temperature, often to assess envelope performance, retrofit strategies, and thermal comfort using energy balance equations. Tools such as EnergyPlus and TRNSYS dominate these applications due to their ability to integrate measured boundary data, detailed material properties, and realistic occupancy patterns. For instance, Petrou et al. [45,46] used EnergyPlus to compare overheating predictions across residential prototypes, while Li et al. [47] and Kang et al. [48] applied nodal RC networks to evaluate passive solar design and dynamic thermal responses in energy-efficient envelopes. Analytical RC methods, such as those by Kang et al. [48], provided transparent, reduced-order formulations for sensitivity studies and control optimization, demonstrating the adaptability of nodal modeling across contexts.
Fig. 4.
Counts of Nodal Model Tools to the Predicted Variables as Used in the Reviewed Studies.1
Recent research used nodal models to include coupled heat and moisture transfer, enabling simultaneous prediction of indoor temperature and relative humidity. Studies by Li et al. [42], Garcia-Frometa et al. [49], and Gamboa-Loya et al. [50] implemented EnergyPlus-based nodal frameworks to capture hygrothermal behavior under varying wall systems and climates, revealing how envelope moisture storage capacity influences indoor humidity regulation. Choi et al. [51] compared simplified and detailed moisture balance models, showing that parameterization of sorption and diffusion processes can substantially affect RH predictions. These advances illustrate the growing ability of nodal frameworks to couple thermal and hygric behavior efficiently while maintaining manageable computational demand.
A smaller group of studies focused primarily on humidity-related performance, though temperature was still internally resolved within the nodal heat balance. Winkler et al. [41] examined how indoor RH affects occupant comfort and cooling energy demand using EnergyPlus simulations, while Martin et al. [35] developed a lumped moisture-balance model for internal moisture generation calibrated against field measurements in occupied homes. These studies demonstrate that nodal approaches can address humidity as a key outcome variable while preserving thermodynamic consistency.
In terms of predictive performance, nodal models have shown strong agreement with experimental and field data across residential applications. Rincón et al. [52] and Liu et al. [28] achieved mean indoor temperature deviations below 1 °C and RH errors within ±2 % when simulating traditional and high-mass wall constructions, respectively. Xu et al. [53] validated PCM retrofits in cold-climate apartments with CV(RMSE) below 12 %, while Santamaría et al. [54] and Gamboa-Loya et al. [50] replicated dynamic temperature behavior with less than 1.5 °C RMSE. Even in humidity-focused studies, such as Tang et al. [55], EnergyPlus nodal solvers captured relative humidity bias within 5 %, confirming their reliability when key physical parameters are accurately defined.
As illustrated in Table 2, nodal modeling has been extensively applied across single-family and multifamily buildings, were inter-zone airflows and thermal coupling play significant roles. Input parameters typically include zone-to-zone flow coefficients, infiltration paths, window and vent opening schedules, internal gains, and material thermal properties to simulate built environment settings and performances. For instance, Zhao et al. [56] used a TRNSYS–MATLAB optimization framework to assess insulation and ventilation performance in prefabricated dwellings, while Grygierek et al. [57] leveraged multi-zone simulations to design natural and fan-assisted cooling strategies across insulation types.
Table 2.
Nodal model tools applied across residential building typologies.
| Tool | EnergyPlus | Analytical | TRNSYS | Others | Total |
|---|---|---|---|---|---|
| Single-family house | 9 | 1 | 3 | – | 13 |
| Multi-family / apartment | 4 | – | – | 1 | 5 |
| Prototype / test cell | 2 | 2 | 1 | 1 | 6 |
| High-rise apartment | – | – | – | – | – |
| Passive / zero-energy (SF) | – | 1 | – | – | 1 |
| Historic / traditional | 1 | – | – | – | 1 |
| Prefab / modular (SF) | – | – | – | – | – |
| Total | 16 | 4 | 4 | 2 | 26 |
Overall, nodal models balance physical transparency, computational efficiency, and strong empirical validation. Their proven accuracy across diverse residential typologies, ranging from traditional dwellings to modern high-performance buildings highlights their robustness in predicting both temperature and humidity dynamics. While the assumption of well-mixed air limits their representation of stratified or localized microclimates, their consistent ability to capture integrated hygrothermal performance reinforces their status as a foundational physics-based approach for evaluating residential indoor environments across building types and climates.
4.1.2. Zonal models
Zonal models represent an intermediate level of detail between lumped nodal and high-resolution CFD approaches by dividing indoor spaces into interconnected subzones governed by mass, momentum, and energy conservation. Each zone exchanges heat and air with adjacent volumes, allowing the model to capture vertical temperature gradients, airflow asymmetry, and localized comfort variations while maintaining moderate computational cost. As shown in Fig. 5, zonal methods have been widely applied to analyze indoor temperature distribution, natural ventilation, and thermal comfort performance in residential buildings across different climates and construction types.
Fig. 5.
Counts of Zonal White Box Modeling Tools to the Predicted Variables.
As shown in Fig. 5, most zonal studies primarily focused on temperature prediction, leveraging the multi-zone solvers in EnergyPlus and TRNSYS. Zhao et al. [56] used TRNSYS to evaluate retrofit options for prefabricated homes in cold regions, finding that enhanced insulation and window optimization reduced indoor temperature fluctuations by 1–2 °C. Eguía-Oller et al. [58] applied TRNSYS–TRNFLOW to replicate the multi-zone dynamics of the IEA Twin House, achieving RMSE values below 1 °C and confirming the model’s capability to reproduce measured temperature profiles. Similarly, Kadri et al. [59] simulated reflective double-skin roofs in hot-arid dwellings using TRNSYS, showing up to 5 °C reductions in operative temperature and over 60 % cooling-energy savings. Radujković et al. [60] used EnergyPlus to analyze modular green-wall retrofits, observing reductions in summer indoor temperatures by up to 3.5 °C and winter heating energy savings around 6 %, with model–measurement deviations under 1.5 °C.
Although the majority of zonal studies emphasize indoor temperature dynamics, several integrate humidity as a secondary or coupled parameter within heat balance simulations. For example, Abbaas et al. [61] incorporated RH tracking into multi-zone EnergyPlus simulations of naturally ventilated buildings, showing that nighttime airflow reduced indoor temperature by up to 10 °C while keeping RH within acceptable comfort ranges.
As illustrated in Fig. 5 and detailed in Appendix A, zonal studies overwhelmingly relied on EnergyPlus and TRNSYS due to their AirflowNetwork and TRNFLOW capabilities, which explicitly model buoyancy-driven and wind-induced airflows between thermal zones. Input variables typically included envelope U-values, emissivity, infiltration coefficients, internal gains, window and door ventilation rates, and outdoor boundary conditions. Firląg et al. [62] employed a three-zone TRNSYS model to capture temperature stratification under future climate scenarios, while Tian et al. [63] used a two-zone EnergyPlus setup to investigate the impact of solar collector retrofits on indoor comfort and heating demand.
Model performance across zonal studies consistently meets high predictive standards when validated against empirical data. Calibrated applications generally reported temperature RMSE below 1 °C and CV (RMSE) under 5 %, aligning with ASHRAE Guideline 14 requirements. Eguía-Oller et al. [58] demonstrated accurate dynamic responses under variable ventilation conditions, while Kadri et al. [59] and Radujkovíc et al. [60] achieved strong agreement between simulated and monitored temperatures. Even in mixed-mode environments, zonal models reliably captured transient effects of natural ventilation and occupant-driven changes.
Overall, zonal modeling provides a physically grounded yet computationally efficient framework for studying interzonal heat transfer, airflow, and coupled thermal–moisture interactions in residential buildings. Its ability to simulate spatial variability, often missed by nodal approaches makes it particularly suitable for assessing ventilation-driven comfort, passive design, and retrofit strategies. Although humidity-only analyses remain rare, zonal models have demonstrated stable performance in reproducing both temperature and relative humidity trends, reaffirming their value as an effective and scalable approach for multi-zone indoor environment prediction. Fig. 6
Fig. 6.
Counts of Building Typologies to the Zonal White Box Modeling Tools Adopted in the Reviewed Studies.
4.1.3. CFD models
CFD models represent the most detailed category of physics-based simulations for predicting indoor temperature, humidity, and airflow behavior. Unlike nodal or zonal approaches, which assume uniform or piecewise-uniform air conditions, CFD models solve the full Navier–-Stokes equations for mass, momentum, and energy conservation within discretized spatial grids. This allows them to capture three-dimensional gradients in air velocity, temperature, and moisture distribution, making them particularly effective for analyzing localized comfort, ventilation effectiveness, and design optimization in residential buildings. CFD applications in indoor environmental prediction have become increasingly diverse, employing solvers such as ANSYS Fluent, OpenFOAM, and DesignBuilder’s integrated CFD module, each varying in turbulence modeling complexity and coupling with thermal boundary conditions.
Most CFD-based residential studies emphasize temperature prediction and airflow characterization under various passive and hybrid ventilation configurations. For example, Leng et al. [64] examined the influence of air-well geometry on indoor temperature and airflow patterns in Malaysian terraced houses using DesignBuilder’s k–ε turbulence model, achieving mean temperature prediction errors of approximately 15 % relative to field data. Zhong et al. [65] applied steady RANS simulations to assess airflow and surface convection in vernacular buildings, finding that the SST k–ω turbulence model provided better agreement with wind tunnel measurements than standard wall functions. These examples highlight CFD’s ability to resolve localized convection and geometric influences that simpler models cannot capture.
A number of studies extended CFD modeling to coupled heat and moisture prediction, demonstrating its applicability for indoor humidity and microclimate assessment. Bonello et al. [66,67] series of studies on high-occupancy spaces employed transient CFD with LRN k–ε and SST k–ω turbulence models to predict specific humidity and relative humidity distribution, achieving accuracy within ±0.0015 g/g of experimental data. These findings revealed significant spatial RH variations up to 31 % across occupant zones, underscoring the necessity of high-resolution airflow modeling for evaluating comfort in dense indoor environments.
Validation outcomes consistently demonstrate that CFD models can achieve high predictive fidelity when properly calibrated with boundary and material data. Belpoliti et al. [68] validated DesignBuilder CFD simulations against 10-day monitored data in an occupied dwelling, finding mean temperature deviations of less than 1 °C across natural ventilation scenarios. In Bonello et al.’s [67] humidity microclimate study, CFD-predicted specific humidity remained within 5 % of experimental reference values, outperforming simpler nodal or zonal approaches in capturing spatial moisture dynamics. These quantitative results confirm CFD’s strength in resolving transient phenomena such as stratification, occupant heat release, and moisture dispersion, which are critical in assessing occupant comfort and ventilation design effectiveness.
CFD tools have been applied primarily to single-room or smallapartment configurations due to their computational demands. Typical input variables include air supply velocity, inlet temperature and humidity, material emissivity, solar gains, and occupant heat sources. Studies such as those by Belpoliti et al. [68] exemplify how CFD outputs (localized air temperature, velocity vectors, and moisture concentration) are validated against field measurements to establish their reliability for design optimization and comfort assessment.
Overall, CFD modeling offers the most spatially explicit and physically rigorous representation of indoor environmental conditions among the reviewed approaches. Its capacity to reproduce detailed airflow and humidity gradients makes it indispensable for evaluating ventilation design, localized overheating, and moisture accumulation risks in residential settings. However, its computational intensity and sensitivity to boundary conditions often limit its use to room-scale analyses or hybrid integrations with faster building performance simulation tools. Despite these constraints, CFD remains the reference standard for physically accurate indoor thermal and moisture analysis, serving as both a validation framework and a diagnostic tool for refining lower-order nodal and zonal models.
4.1.4. Hybrid physics based models
Hybrid physics-based models represent an advanced evolution of white-box simulation, combining multiple modeling paradigms such as nodal, zonal, and CFD-based approaches with whole-building energy simulation (BES) tools to capture coupled thermal, airflow, and hygrothermal behavior. These models balance physical detail and computational efficiency by linking solvers across different spatial and temporal domains. As shown in Fig. 7, hybrid frameworks identified in this review fall into four major categories: CFD and BES, EnergyPlus-based hybrids, TRNSYS-based hybrids, and multi-platform co-simulation models. Cross-referencing with Fig. 8 shows that hybrid models are most frequently applied to single-family and multi-family dwellings, followed by high-rise apartments and prototype or test cell buildings, reflecting their adaptability across different residential typologies.
Fig. 7.
Counts of the Hybrid Physics Based Modeling Tools to the Predicted Variable in the Reviewed Studies.2
Fig. 8.
Counts of the Studies Using Hybrid Physics Based Modeling Types Stratified by the Building Types.
CFD and BES hybrids provide detailed spatial resolution of indoor airflow and temperature distribution while remaining less computationally intensive than standalone CFD simulations. These hybrids integrate CFD solvers with energy models to capture buoyancy, stratification, and convection effects in naturally ventilated or free-running buildings. Serageldin et al. [69] coupled ANSYS Fluent with TRNSYS to evaluate a solar chimney and earth tube cooling system in Egyptian homes using input variables such as air velocity, turbulence intensity, surface heat flux, and solar radiation. The hybrid achieved temperature reductions of 5–9 ° C, validating its capability to simulate convective cooling under high outdoor temperatures. Ayoobi et al. [70] used DesignBuilder’s CFD module to assess windcatcher performance in Kabul, applying wind pressure coefficients, convective heat transfer coefficients, and envelope emissivity as key parameters. The study showed that natural cross-ventilation could maintain indoor comfort within ASHRAE 55 standards in arid climates. These examples demonstrate that CFD+BES hybrids are particularly suited for single-family houses, where airflow, orientation, and passive cooling strategies strongly affect thermal comfort.
EnergyPlus-based hybrids represent the most widely used configuration due to EnergyPlus’s capability to integrate airflow and moisture models such as the AirflowNetwork (AFN), Heat and Moisture Transfer (HAMT), and co-simulation with CONTAM. Gao et al. [71] applied the AFN module to examine stack and wind-driven ventilation in Hong Kong high-rise buildings using window opening fractions, infiltration rates, and inter-zonal pressure coefficients as primary inputs. The model showed a 20 percent increase in natural ventilation effectiveness and improved thermal comfort hours. Belloum et al. [72] used the HAMT module to simulate moisture buffering in palm concrete test cells, incorporating vapor permeability, sorption isotherms, and material porosity. The study achieved relative humidity (RH) errors below 2.4 percent, showing EnergyPlus’s effectiveness in modeling hygrothermal coupling. Grygierek et al. [57] expanded this through an EnergyPlus–CONTAM co-simulation, linking infiltration coefficients, crack-flow parameters, and CO2 control setpoints to evaluate indoor air quality (IAQ) and comfort. These hybrids predominantly predicted temperature and RH, achieving errors below 1 ° C and 3 percent RH, and were most frequently applied in multi-family and single-family buildings.
TRNSYS-based hybrids excel in transient, high-resolution studies, especially for experimental or test cell research. Ju et al. [73] coupled TRNSYS Types 399 and 1270 to model PCM-integrated radiant floors in rural Chinese homes, using enthalpy-temperature relationships, slab thickness, and convective coefficients as inputs, achieving temperature RMSE under 0.8 °C. Belloum et al. [72] used TRNSYS Type 56 to calibrate moisture-buffering in a bio-based test cell through air-change rates, moisture capacity, and vapor permeability. Martínez-Marino et al. [74] combined TRNSYS with HAMFitPlus to simulate multi-zone indoor conditions in social housing, achieving temperature RMSE under 0.5 °C and RH deviation under 3.6 percent. These studies confirm that TRNSYS hybrids deliver high temporal accuracy and flexible calibration, making them ideal for prototype and experimental buildings requiring detailed hygrothermal assessment.
Multi-platform and co-simulation hybrids integrate specialized tools to represent complex interdependencies among temperature, humidity, and air quality. Heibati et al. [75] coupled EnergyPlus, CONTAM, and WUFI to assess indoor moisture and IAQ in a Canadian home using air permeability, latent load rates, and vapor diffusion coefficients. The model achieved strong alignment across RH and enthalpy predictions. Van Hove et al. [76] applied Modelica–Dymola for a zero-energy home, linking thermal, RH, and CO2 control variables through heat transfer coefficients, HVAC signals, and diffusion constants, maintaining temperature deviations below 1 °C and RH within 5 percent. Though computationally demanding, such hybrid frameworks deliver exceptional performance accuracy and enable real-time analysis of integrated building systems.
Across all reviewed studies, hybrid physics-based models consistently achieve temperature errors below 1 °C and RH deviations within 5 percent, confirming their superior predictive accuracy compared with single-domain models. The predicted variables most frequently include temperature and relative humidity, and then humidity-only analyses, focused on moisture risk or buffering effects. In terms of building typology, hybrids are primarily applied to single-family and multi-family buildings, while test cell and prototype studies rely on TRNSYS configurations for precise experimental replication. Common input parameters include thermal conductivity, vapor diffusion resistance, infiltration rates, material moisture capacity, and air exchange coefficients, each crucial for ensuring consistent heat and mass transfer coupling. As illustrated in Figs. 7 and 8 and detailed further in Appendix A, hybrid models stand out as the most comprehensive and physically grounded simulation framework for reproducing temperature and humidity dynamics in residential buildings, offering a balance between computational practicality and multi-domain precision.
4.2. Trend of white box modeling techniques
In earlier years (2019–2020) as shown in Fig. 9, nodal models represented the most prevalent approach. Of the 13 physics-based studies published in 2019, eight were nodal, reflecting traditional preferences due to their ease of implementation, moderate computational demand, and sufficient accuracy for large-scale energy assessments or long-term scenario analyses. During this period, typical applications involved annual energy performance, envelope retrofit assessments, and broad thermal comfort predictions. Exemplary studies include Petrou et al. [45], comparing overheating predictions with EnergyPlus versus IES VE in UK residential flats, and Li et al. [42], simulating thermal behaviors of rural housing in China. The presence of CFD (two studies in 2019) and hybrid physics-based models (two studies in 2019) was limited.
Fig. 9.
Trend of the White Box Modeling Approaches.
A distinct methodological shift occurred around 2021–2022, marked by increased complexity and coupling of modeling approaches. Despite a slight reduction in overall studies from 2020 to 2021, hybrid physics-based modeling emerged as the dominant approach in 2021, with seven of ten studies embracing multiple coupled domains (thermal, airflow, moisture). Notable examples included integrated simulation frameworks such as Heibati et al. [75], which combined EnergyPlus, CONTAM, and WUFI to assess indoor climates, and Martínez-Marino et al.’s [74] multi-objective calibration linking TRNSYS with TRNFLOW and HAMFitPlus to rigorously quantify moisture-buffering capacities in residential buildings. Conversely, nodal, zonal, and CFD categories each accounted for only a single study in 2021, emphasizing the shift toward integrative modeling approaches to better represent complex interactions between building physics domains. In 2022, hybrid physics-based studies remained prominent (five out of eight studies), reinforcing this emerging trend. Concurrently, zonal models saw increased adoption (two studies), serving as efficient alternatives when airflow patterns were important but did not demand full CFD analysis. Nodal modeling dropped noticeably, represented by only one publication, signifying a community transition toward more spatially resolved and physically integrated simulations.
The year 2023 witnessed a resurgence of zonal modeling (five out of nine studies). Researchers employed zonal approaches primarily through tools such as EnergyPlus AirflowNetwork and TRNFLOW to balance spatial resolution and computational efficiency, effectively capturing inter-zonal airflow and temperature gradients without the complexity of full CFD. For instance, Yeretzian et al. [77] utilized EnergyPlus AirflowNetwork models to optimize hybrid ventilation strategies in Beirut apartments, and studies like those conducted by Eguía-Oller et al. [58] extended zonal methods to the German Twin House to explore advanced ventilation controls.
By 2024, the highest annual total (15 studies) indicated peak research activity across all modeling types, likely driven by heightened global interest in indoor air thermal conditions and resilience against climatic extremes post-pandemic. Hybrid physics-based and zonal models dominated, each representing five studies. Hybrid approaches continued addressing comprehensive building physics interactions, exemplified by studies like Abbaas et al. [61], integrating thermal and airflow models to optimize building forms for natural ventilation in Jordan. Nodal models re-emerged slightly (four studies), but with more sophisticated applications, often incorporating uncertainty analysis, predictive control, or Bayesian calibration, as illustrated by Petrou et al.’s [46] archetype calibration in UK housing. CFD maintained a minor yet significant presence with detailed analyses like Belpoliti et al.’s [68] CFD-based validation of natural ventilation performance in UAE residential dwellings.
The apparent decline in studies observed in 2025 reflects the early cutoff date than actual reduced interest. Nonetheless, nodal methods persisted as the most represented (three studies), such as Xu et al.’s [53] PCM retrofit analysis for cold-climate apartments in Harbin, China, and Zhao et al.’s [56] optimization of insulation strategies using TRNSYS and genetic algorithms. Zonal modeling-maintained relevance with a single study but no further CFD or hybrid studies yet reported in this early stage.
Across the full seven-year period, nodal models accounted for 26 of the total 71 physics-based studies, underscoring their continued role as accessible baseline simulation approaches. Hybrid physics-based methods closely followed with 24 studies, highlighting their growing importance in tackling integrated heat-air-moisture problems. Zonal models (16 studies) gained substantial momentum, especially in recent years, due to their balance between detail and computational practicality. CFD remained specialized, with only five studies total, reflecting its targeted use primarily in detailed airflow, pollutant dispersion, or localized thermal-comfort analyses.
4.3. Limitations of white box modeling techniques
While white box modeling approaches, such as nodal, zonal, CFD, and hybrid physics-based models, have advanced the predictive understanding of indoor thermal environments in residential buildings, each method comes with specific limitations that affect its applicability, performance, and computational cost. To facilitate a structured comparison, Table 3 summarizes the core constraints associated with each technique across six critical modeling dimensions: computational efficiency, prediction accuracy, humidity representation, boundary assumptions, occupant/HVAC integration, validation practice, and scalability to real-world buildings.
Table 3.
Cross-Technique Comparison of White box Model Limitations.
| Aspect of Limitation | Nodal Models | Zonal Models | CFD Models | Hybrid Physics-Based Models |
|---|---|---|---|---|
| Computational Efficiency | Very efficient and suitable for rapid prototyping [48], but oversimplifies airflow and moisture transfer | Moderate efficiency, scales with number of zones; faster than CFD but less granular [46] | Poor efficiency; meshing and solver demands lead to long runtimes [66] | Varies; coupling can improve efficiency but often increases complexity [74] |
| Prediction Accuracy | Good for mean temperature, but weak in capturing spatial gradients or dynamic RH [56] | Captures stratification better than nodal, but only approximate [80] | High fidelity for airflow, stratification, and RH gradients [66,67] | Generally improves accuracy by combining methods, but calibration-dependent [74] |
| Humidity Representation | Capable of simulating humidity when coupled with hygrothermal modules, but frequently omitted in practice due to simplified setups or unavailable material data [46,50]. | Can represent spatial RH variations if moisture transport equations are defined; most implementations remain temperature-focused for computational simplicity [74] | Explicit RH fields solvable through coupled mass-transfer equations but computationally expensive for long-term simulation [66,67]. | Commonly include hygr other mal or air-moisture coupling, allowing more realistic humidity prediction but requiring detailed calibration and input data [72,75]. |
| Boundary Assumptions | Simplified, assuming well-mixed zones and steady surface properties [43] | Often static inlets/outlets and fixed zone boundaries [37,57] | Requires detailed specification of outdoor wind, turbulence, and wall functions [68,77] | Boundary realism improves via coupled BES-CFD or hygrothermal models, but adds input/data burden [70,74]. |
| Occupant/HVAC Dynamics | Typically idealized with fixed occupancy or HVAC schedules [46,48] | Can incorporate occupant behavior but often static or oversimplified [34] | Rarely integrated, unless co-simulated with BES [68] | More flexible for integrating realistic occupancy and HVAC interactions, especially when linking EnergyPlus/TRNSYS with airflow or moisture tools [57,75] |
| Validation Practice | Frequently validated with limited field data or short campaigns [56] | Often validated against monitored apartments or archetypes but with seasonal scope [81] | Validations limited to lab-scale or short-term monitoring; cosdy for real homes [66,68] | Typically benchmarked with calibrated experimental houses or multi-tool comparisons T741 |
| Scalability to Real Homes | Easily scaled across archetypes and climates, though with low physical granularity [43,55] | Suitable for multizone and archetype studies [34,46] | Difficult to scale due to high runtime and setup demands [65,66,68] | L/ tJ Scalable via modular integration of BES, CFD, and hygrothermal components, but requires expertise and calibration [74] |
Nodal models, though computationally lightweight and well-suited for large-scale or control-oriented applications, tend to oversimplify spatial dynamics and neglect transient interactions such as moisture buffering or stratification. Their limited boundary realism and often static HVAC assumptions reduce their performance in heterogeneous indoor environments. Zonal models improve upon this by dividing space into multiple volumes, enabling stratification and directional airflows. However, they require increased computational resources and still rely on simplified inlet/outlet conditions, making them sensitive to the number of zones and the assumptions behind their separation. Neglecting the moisture-buffering behavior of building materials represents a major limitation rather than a minor simplification. Hygroscopic surfaces such as wood, plaster, or bio-based composites actively absorb and release water vapor, moderating indoor humidity fluctuations and stabilizing perceived comfort levels. When these effects are omitted, models tend to overpredict short-term humidity swings, underestimate latent loads, and misrepresent the interaction between envelope materials, ventilation, and thermal comfort. This can lead to deviations of up to 20–35 % in predicted relative humidity and associated comfort indices, particularly in lightweight or naturally ventilated dwellings. Accounting for buffering behavior is therefore essential for realistic simulation of indoor comfort and energy performance, especially in humid and mixed-humid climates [78,79].
CFD models offer the most detailed insight into airflow and temperature distribution, capturing convection, buoyancy, and microclimatic gradients with high spatial resolution. Nonetheless, their high computational burden, demand for finely meshed geometries, and complex boundary input requirements limit their scalability in multiunit or time-intensive simulations. Furthermore, full-scale CFD validation remains rare due to the logistical and sensor density challenges of replicating real airflow conditions in situ.
Hybrid physics-based models aim to integrate the strengths of multiple white box paradigms. By coupling zonal or nodal cores with detailed hygrothermal or airflow modules (or with BES tools like EnergyPlus), these models achieve higher realism and versatility. However, they require careful calibration and are often dependent on cosimulation strategies that can introduce their own sources of numerical instability or overhead. Their performance tends to vary depending on the integration depth and the quality of boundary or climate data inputs.
5. Black box modeling
Black box models, often referred to as data-driven models, have emerged as a powerful alternative to physics-based (white box) approaches for predicting indoor thermal conditions. Rather than explicitly simulating the underlying thermophysical processes of buildings, these models employ machine learning and statistical techniques to uncover patterns, correlations, and nonlinear relationships directly from historical and real-time data [82]. Their strength lies in the ability to bypass detailed construction and material parameters, which are often unavailable or prohibitively expensive to obtain, and instead leverage sensor streams, weather data, and operational records.
In contrast to white box models, which are rooted in fundamental physical laws, black box approaches treat the building as a system whose outputs (e.g., indoor temperature or humidity) can be predicted from measured inputs (e.g., outdoor conditions, occupant behavior, HVAC operation). This makes them particularly attractive for large-scale residential datasets, smart thermostat applications, and cases where rapid or transferable predictions are required. For example, long short-term memory (LSTM) neural networks have been shown to anticipate indoor temperatures with sub-1 °C error in naturally ventilated dwellings [83], while transfer learning methods have reduced commissioning data requirements by nearly 90 % for predictive building control [84].
The versatility of black box modeling is evident in its wide range of techniques, spanning from shallow and deep neural networks, regression-based models, and hybrid/ensemble frameworks. Each comes with its own balance of accuracy, interpretability, and computational demand. Recent studies demonstrate their potential not only for temperature forecasting but also for multi-variable prediction, including RH, comfort indices, and even energy savings linked to indoor climate control. This breadth of applications underscores the growing importance of black box models as scalable tools for indoor environmental performance analysis, especially as sensor-rich residential datasets continue to expand. For broader details on individual studies, kindly refer to Appendix Table B.
5.1. Model types: performance, predicted variables, and building types
5.1.1. Shallow neural networks
Shallow Neural Networks (SNNs), typically represented by single-layer Feed-Forward Neural Networks (FFNNs) and simple Multilayer Perceptrons (MLPs), are among the earliest black box approaches applied to predict indoor conditions in residential buildings. Their relatively simple architecture makes them computationally efficient and suitable for moderate-sized datasets, providing a practical entry point for data-driven prediction.
As shown in Fig. 10, approximately 60 % of SNN applications focused exclusively on indoor temperature, 10 % on humidity only, and the remainder jointly modeled both. This reflects the ready availability of temperature data and its direct links to occupant comfort . In terms of evaluation approaches, Fig. 11 indicates that around half of SNN studies relied on quantitative error measures such as RMSE and MAE, while about 22 % employed correlation-based measures, and fewer used classification or descriptive comparisons.
Fig. 10.
Counts of the Predicted Variables by Black Box Model Types Used in the Reviewed Studies.
Fig. 11.
Counts of the Performance Metrics Used Across the Black Box Model Types in the Studies.
The reviewed SNN studies predominantly utilized simple environmental and operational inputs, such as outdoor dry-bulb temperature, solar radiation, dew-point temperature, wind speed, and HVAC energy use. Sözer et al. [40] trained a three-layer MLP using seven inputs (outdoor dry-bulb and dew-point temperatures, wind speed and direction, solar azimuth, atmospheric pressure, and simulated heating energy) and achieved prediction errors under 1.3 °C for temperature and 2 % for relative humidity. Potočnik et al. [85] further integrated TRNSYS-simulated and measured weather data (including solar radiation and outdoor temperature) into their neural and ELM frameworks, reducing RMSE by 35 % relative to linear ARX models. These examples underscore how even shallow architectures perform reliably when equipped with mixed indoor–outdoor and energy-related variables.
The distribution of building types using SNNs (Fig. 12) shows the largest share in single-family homes (42 %), followed by multi-family apartments (25 %), and dormitories or elderly-care facilities (17 %). Their prevalence in smaller-scale and structured residential contexts reflects both the simplicity of the models and the homogeneity of datasets typically available. Nonetheless, limitations persist, shallow networks struggle to capture long-term temporal dependencies or the nonlinear interactions arising from occupancy, ventilation, and thermal inertia. These shortcomings restrict their generalizability across climates and building archetypes, motivating the shift toward deeper or hybrid models.
Fig. 12.
Counts of Building Types by the Black Box Model Types.
5.1.2. Deep neural networks
Deep Neural Networks (DNNs) extend the capacity of shallow networks by incorporating multiple hidden layers and advanced temporal architectures such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRUs), and attention mechanisms. These models are particularly effective at capturing long-term dependencies and complex nonlinearities in indoor climate dynamics, making them highly relevant for predictive control applications in residential buildings.
In terms of predicted variables, Fig. 10 shows that about 40 % of DNN applications targeted both temperature and humidity, compared to 60 % that predicted temperature alone, demonstrating greater versatility than SNNs. Fig. 11 illustrates a balanced use of evaluation strategies, with roughly 40 % of studies reporting error-based measures (e.g., RMSE, MAE), 30 % correlation metrics, and the remainder split between classification and qualitative assessments.
Most DNN-based studies used a broader and more dynamic set of input variables than SNNs, combining meteorological parameters (outdoor temperature, humidity, solar radiation, and wind) with operational or occupant-related data. For instance, Weng et al. [83] trained an LSTM-RNN on simulation-based datasets incorporating window-opening schedules and natural ventilation states, achieving over 93 % of predictions within 2 °C of measurements. Song et al. [86] employed a hierarchical attention GRU model (HAGRU) using multi-zone indoor temperatures, outdoor conditions, and heating loads as inputs, reaching above 98 % accuracy. Chen et al. [84] advanced this approach with CNN-LSTM transfer learning by fine-tuning models using only 15 days of local data enriched with outdoor temperature, humidity, and HVAC operation, significantly improving generalization.
Other DNNs also expanded input diversity: Alhamayani et al. [31] leveraged smart Wi-Fi thermostat data combined with outdoor temperature and solar heat gain estimation; Boubouh et al. [87] optimized temporal abstraction using LSTM time sequences; and Qi et al.’s [88] CONST network used spatio-temporal consistency modules to enhance dynamic feature learning. This broad input integration demonstrates that DNNs excel when trained on multi-source, high-frequency datasets that reflect both environmental forcing and human–system interactions.
The building-type distribution (Fig. 12) highlights that DNNs have been used broadly across single-family homes (≈36 %), apartments (≈27 %), and mixed-use residential settings (≈27 %). This balance reflects their adaptability to diverse data contexts, from controlled test cases to real-world multi-family dwellings. Despite their strong performance, DNNs are data- and computation-intensive, requiring careful hyperparameter tuning and often suffering from reduced interpretability. However, techniques such as attention mechanisms and explainability frameworks are increasingly being integrated to address these issues.
5.1.3. Regression-Based models
Regression-based models, including classical linear regression, regularized forms such as ridge and lasso, and auto-regressive models with exogenous variables (ARX), remain a cornerstone of black box modeling in residential settings. Their advantages are transparency, computational efficiency, and robustness when applied to large datasets or short-term forecasting tasks, even if they lack the representational power of neural networks.
As indicated in Fig. 10, the majority of regression studies (≈65 %) focused solely on indoor temperature prediction, with fewer addressing humidity or combined variables. Fig. 11 shows that nearly 55 % of regression-based evaluations used error metrics, 25 % correlation measures, and very few relied on classification or qualitative assessment.
Most regression-based models relied on measured environmental and operational inputs, particularly outdoor dry-bulb temperature, solar radiation, heating energy use, and occupancy or thermostat schedules. Al-Obeidat et al. [89] trained ridge and lasso regressions using historical smart-thermostat sensor data combined with outdoor temperature and solar heat input, integrating Fanger’s PMV model to evaluate thermal comfort thresholds. Huchuk et al. [30] extended similar models using aggregated thermostat data from over 1000 homes across North America, incorporating time-of-day, outdoor weather, and setpoint histories. Benzaama et al. [90] employed a piecewise auto-regressive exogenous (PWARX) model trained on dormitory monitoring data, where key exogenous variables included heating power and solar radiation. Pergantis et al. [91] advanced regression further by introducing a humidity-aware model predictive control framework that used sensible heat ratio and latent load as dynamic inputs to optimize HVAC operation. Collectively, these studies highlight that regression-based models perform best when supplied with physically meaningful predictors directly linked to energy balance and indoor dynamics.
The building-type distribution (Fig. 12) shows regression models most frequently applied in single-family houses (≈50 %), with smaller shares in multi-family apartments, dormitories, and mixed-use contexts. This concentration reflects the suitability of regression for straightforward, well-instrumented cases. However, regression’s linear structure limits its ability to capture nonlinear effects and cross-variable interactions, placing it behind neural and ensemble approaches in terms of accuracy and adaptability.
5.1.4. Hybrid and ensemble models
Hybrid and ensemble models combine the strengths of multiple learners or modeling paradigms to enhance predictive power, robustness, and generalization. Hybrids may integrate different machine learning architectures (e.g., CNN-LSTM-AE) or combine statistical and data-driven approaches, while ensembles typically aggregate predictions across heterogeneous models.
Fig. 10 shows that nearly one-third of hybrid and ensemble studies simultaneously predicted both temperature and humidity, a larger share than in other black box categories. Fig. 11 indicates that in addition to error and correlation metrics, about 20 % of these studies explicitly reported hybrid-specific evaluation methods, underscoring their comparative role.
The reviewed hybrid and ensemble models demonstrated high diversity in their input configurations, often combining meteorological, occupancy, HVAC, and air-quality variables. Yu et al. [92] developed a deep-ensemble model incorporating Random Forest (RF), Support Vector Machine (SVM), LSTM, and XGBoost, using outdoor temperature, humidity, solar radiation, and multi-zone indoor sensor data to achieve R2 values above 0.96 across 25 residential units in Australia. Jogunola et al. [93] employed a CNN–Bidirectional LSTM–Autoencoder framework that used eight meteorological inputs (temperature, humidity, wind speed, and dew point) along with time and calendar features to enhance temporal consistency. Espinosa et al. [32] integrated Random Forest, LSTM, and SVM surrogates within a multi-objective evolutionary algorithm, relying on multivariate time-series features such as indoor temperature, solar radiation, and occupancy-driven schedules. Sung et al. [94] combined CNN and LSTM models with per-minute indoor sensor streams and outdoor meteorological data to simultaneously predict temperature, humidity, and CO2 concentration, with per-minute RMSE as low as 0.042 °C for temperature. Together, these hybrid configurations reveal that coupling multi-sensor indoor data with external climate and control variables significantly enhances model robustness and transferability across building types.
Building-type distribution (Fig. 12) shows hybrids and ensembles deployed across a broad spectrum: 31 % in single-family homes, 25 % in apartments, and 25 % in mixed-use or smart home environments. Their ability to operate effectively across diverse residential contexts highlights their flexibility. Despite clear gains in accuracy and robustness, hybrids and ensembles are computationally demanding, may require extensive tuning, and risk overfitting if not carefully validated. Nonetheless, their rising prominence reflects a growing emphasis on integrated, multi-variable, and scalable modeling approaches for residential buildings.
Taken together, the synthesis of Figs. 10–12 and supporting studies from Appendix Table B demonstrates a clear progression in the black box modeling landscape. Shallow and regression-based models continue to serve as computationally efficient baselines but are largely limited to single-variable (temperature-only) applications and narrower residential contexts. Deep neural networks significantly improve accuracy, especially for multi-variable prediction, though they demand higher data volumes and computational power. Hybrid and ensemble approaches, however, provide the most balanced solution: they yield consistent sub-1 °C errors, generalize across climates and building types, and integrate both temperature and humidity prediction. By spanning nearly all metric categories, hybrid models reflect both methodological rigor and practical utility. Thus, while no single model type can be universally prescribed, the evidence strongly suggests that ensemble and hybrid frameworks will dominate future black box applications in residential indoor climate prediction, offering the best trade-off between scalability, interpretability, and performance.
5.2. Trend of black box modeling techniques
The temporal distribution of black box modeling studies reveals how methodological preferences have shifted between 2019 and 2025 (Fig. 13). Early work in 2019 and 2020 was characterized by a balanced presence of shallow neural networks, regression-based models, and deep learning. For example, Weng et al. [83] applied recurrent neural networks to naturally ventilated housing in the UK, while Potočnik et al. [85] benchmarked neural networks against linear ARX models. During this same period, regularized regression techniques such as ridge and lasso were adopted for smart-home forecasting [89].
Fig. 13.
Evolution of the Black Box Model Types.
Between 2021 and 2023, the share of hybrid and ensemble methods grew steadily, representing close to 40 % of the total studies by 2023. This reflects an increasing recognition that single algorithms often underperform when facing highly variable indoor environments. For instance, Yu et al. [92] demonstrated that deep ensemble machine learning maintained RMSE within 0.5 °C across 25 mixed-use buildings in Australia, while Espinosa et al. [32] employed surrogate-assisted evolutionary algorithms to optimise feature selection, achieving around 13 % improvement over wrapper baselines. These advances suggest a broader turn toward integrative models capable of combining predictive accuracy with robustness across diverse datasets.
By 2024 and 2025, the trend solidified further, with hybrid and ensemble approaches becoming the dominant category. Approximately half of all studies published in 2024 employed these methods, such as Hu et al. (2024), who introduced a self-learning dynamic graph neural network that improved R2 by nearly 25 % over standard LSTMs in residential HVAC forecasting. Similarly, Farahani et al. [38] compared gradient boosting ensembles against LSTMs for overheating prediction in Nordic apartments, showing that boosted ensembles better captured daily fluctuations during heatwave periods. At the same time, regression-based methods and shallow neural networks declined sharply, together accounting for fewer than 25 % of studies after 2023.
In summary, Fig. 13 highlights a clear trajectory in black box modeling, as the trend moves away from early reliance on shallow networks and regression toward more complex, hybridized frameworks that integrate multiple algorithms. This reflects a broader paradigm shift from accuracy-focused benchmarking to models designed for generalization, interpretability, and scalability across climate zones and building types. For detailed study-level breakdowns underpinning these patterns, readers are referred to Appendix Table B.
5.3. Limitations of black box modeling techniques
Table 4 highlights that the shortcomings of black box models for indoor climate prediction are not isolated issues but interconnected themes that collectively limit their robustness and real-world utility. The distribution of limitations shows that while these models can achieve strong predictive accuracy in controlled settings, their performance is heavily tied to the conditions under which they are trained and validated.
Table 4.
Limitations of the Black box Modeling Techniques.
| Category of Limitation | Description | Representative Studies & Examples |
|---|---|---|
| Data Dependency & Representativeness | Black box models require large, representative datasets. Short-term monitoring or simulation-derived data reduces robustness. Transferability suffers when training data differ from new environments. | Weng-Mourshed et al. [83] – reliance on simulation data without real occupancy; Sözer et al. [40] – short monitoring period biases ANN accuracy; Potočnik et al. [85] – TRNSYS-only training limits real-world applicability. |
| Generalizability & Transferability | Models often tailored to one site/building, limiting use across climates, seasons, or building types. Transfer-learning approaches require similarity between source and target data. | Yang et al. [95] – calibrated to one Chinese city; Chen et al. [84] – performance drops if source–target climates differ; Alshammari [96] – single-dwelling training; Hu et al. [97] – short lab experiment, single cooling season. |
| Occupancy & Internal Gains Neglect | Many models exclude occupant schedules, lighting, appliances, and activity effects. This limits accuracy in predicting realistic thermal and humidity conditions. | Sözer et al. [40] – internal gains not explicitly modeled; Huchuk et al. [30] – occupancy/activity absent; Lara et al. [82] – unoccupied room; Boubouh et al. [87] – ignores RH and occupancy effects. |
| Humidity & Latent Loads Ignored | Focus remains on temperature; latent heat and RH dynamics are rarely incorporated, reducing value for comfort and IAQ analysis. | Yang et al. [95] – temperature-only; Benzaama et al. – humidity not addressed; Alhamayani et al. [31] – PMV-based but no humidity modeling; Pergantis et al. [91] – humidity dynamics difficult to capture. |
| Interpretability & Transparency | Black box models lack physical interpretability, making it hard to understand or trust predictions. These limits use in design and regulation. | Sözer et al. [40] – limited physical insight; Chen et al. [84] – interpretability limited; Qi et al. [88]– factor selection hard to justify; Shakhovska et al. [98]– accuracy issues with sparse data. |
| Reliance on External Forecasts | Forecast-driven inputs (weather, solar, ventilation) introduce uncertainty; accuracy declines when forecasts are off. | Potočnik et al. [85] – depends on weather forecast quality; Farahani et al. [38] – errors propagate from heat-wave forecast; Kim et al. [99] – extreme weather not fully captured. |
| Error Propagation & Stability | Multi-step predictions or sequential coupling of models amplify errors over time, undermining reliability for long horizons. | Delcroix et al. [39] – sequential ANN models cause error propagation; Farahani et al. [38] – LSTM misses short-term fluctuations; Sung et al. [94] – CNN-LSTM underperformed simple LSTM, error drift risk. |
| High Complexity vs. Practicality | Sophisticated ML architectures often require computationally heavy training, dense sensor inputs, or fine-grained calibration, limiting deployment in real homes. | Jogunola et al. [93] – envelope/internal gains data unavailable; Laukkarinen et al. [44] – wide variation due to dataset dependence; Sung et al. [94] – computational costs on edge devices not quantified; Yüksek et al. [100] – GPR compute load high. |
| Contextual & Experimental Constraints | Studies often rely on narrow experimental setups, atypical occupancy (COVID-19), or controlled lab tests, restricting real-world applicability. | Yu et al. [92] – COVID-19 occupancy bias; Hu et al. [97] – single lab-controlled residential space; Lara et al. – heat-input-only test room; Espinosa et al. [32] – surrogate FS performance still dependent on base models. |
Another trend evident from Table 4 is the imbalance between computational sophistication and practical deployment. Many of the most advanced architectures introduce added complexity without a proportional gain in reliability, raising questions about cost–benefit and scalability in everyday residential use. The table also makes clear that even when black box models outperform simpler baselines, their lack of transparency remains a critical barrier for adoption in design practice or policy settings where interpretability is as important as accuracy.
6. Conclusion
This review synthesized 91 peer-reviewed studies published between 2019 and 2025 that applied white box and black box modeling techniques to predict indoor temperature and humidity in residential buildings. By systematically analyzing the distribution of modeling approaches, predictive variables, performance metrics, and building contexts, the findings provide a consolidated understanding of how these methods have evolved, where they perform most effectively, and what challenges remain. This review is the first of its kind to comprehensively integrate both physics-based and data-driven paradigms for residential indoor climate prediction, emphasizing the coupled simulation of heat and moisture processes and their relevance to comfort, energy, and health outcomes.
6.1. Application of white box and black box models
White box models remain foundational for representing fundamental building physics. Nodal approaches dominated early applications due to their simplicity and efficiency, while more recent work has shifted toward zonal, CFD, and especially hybrid physics-based models that integrate heat, air, and moisture domains. These methods are particularly strong in scenarios requiring mechanistic explanation, retrofit evaluation, or explicit representation of boundary conditions. However, their predictive performance is strongly influenced by the accuracy of input data and boundary conditions, including occupant schedules and envelope hygrothermal properties.
Black box models, in contrast, have gained momentum for their ability to leverage sensor data and predict short- to medium-term dynamics without detailed physical inputs. Techniques such as LSTM, GRU, CNN-LSTM hybrids, and ensemble architectures consistently achieve sub-1 °C temperature errors and under 3 % RH deviation when well-trained, showing strong potential for real-time control and adaptive operation in residential settings. Shallow neural networks have demonstrated practical accuracy for small-scale applications, while deep neural networks, regression frameworks, and hybrid or ensemble methods now form a diverse toolkit that can outperform traditional physical models under certain conditions. This evolution highlights a methodological convergence between data-driven adaptability and physics-informed realism, particularly as sensor networks and transfer learning expand residential datasets.
6.2. Geographical, seasonal, climatic, and temporal distribution
The reviewed studies span a wide range of contexts, with substantial representation from Europe, East Asia, and North America, but relatively fewer applications in Africa and South America. Seasonal emphasis remains skewed toward heating periods, though recent studies have expanded into cooling and mixed-mode ventilation analyses. Moisture-focused research remains concentrated in humid or mixed-humid climates, with limited representation from arid or tropical regions where latent loads and indoor wet-bulb temperatures significantly affect comfort and health.
Temporally, white box models showed renewed interest after 2020 through zonal–CFD integration, while the share of black box hybrid and ensemble methods grew rapidly after 2022, representing nearly half of all studies by 2024. This growth reflects improved computational tools, expanded IoT and smart thermostat datasets, and broader use of open-source residential energy datasets that facilitate validation and scalability across climates.
6.3. Strengths, limitations, and evolution of modeling techniques
Each modeling family presents unique strengths and limitations. White box models excel in interpretability, physical consistency, and scenario testing, but their limitations stem from simplifying assumptions, computational demands, and sensitivity to poorly defined material and moisture parameters. The neglect of moisture buffering, infiltration variability, and transient airflow interactions can lead to deviations in RH prediction, particularly under naturally ventilated or mixed-mode conditions.
Black box models, on the other hand, demonstrate flexibility and adaptability but are often constrained by data scarcity, sensor quality, and limited physical interpretability. Shallow and regression models perform well for short-term or well-instrumented cases, while deep neural networks and hybrid ensembles show superior robustness for long-term and multi-variable prediction. A key trend across both families (white and black box models) is methodological convergence paving the way for next-generation gray-box modeling frameworks.
6.4. Distribution of predictive variables and performance metrics
Temperature remains the dominant variable across both white box and black box categories, with humidity receiving comparatively less attention despite its critical role in comfort and health [101]. Hybrid black box models and advanced DNNs are beginning to address this gap by predicting both variables simultaneously.
Performance evaluation is likewise fragmented. Error metrics such as RMSE and MAE dominate, while correlation measures, qualitative comparisons, and hybrid-specific metrics appear more selectively. The uneven adoption of performance measures makes cross-study comparison difficult and underscores the need for standardized reporting frameworks.
6.5. Synthesis and future work
Evidence from this review highlights a clear shift toward more integrated and adaptive approaches for predicting residential indoor environments. Physics-based models remain vital for exploring fundamental thermal and moisture dynamics, while data-driven models excel in real-time prediction and control. Evidence from this review highlights a clear shift toward more integrated and adaptive approaches for predicting residential indoor environments. Physics-based models remain vital for exploring fundamental thermal and moisture dynamics, while data-driven models excel in real-time prediction and control. To this end, hybrid or grey-box models that combine the physical interpretability of white box approaches with the learning flexibility of black box models represent a critical next aspect in advancing residential indoor environment modeling. These models can capture the complex, nonlinear interactions between envelope performance, occupant behavior, and HVAC operation while retaining physical consistency and generalizability. The balance of accuracy, interpretability, and computational efficiency positions hybrid models as a key bridge toward fully integrated digital twins and predictive control frameworks for residential buildings. Moving forward, a structured reviewed research agenda can help guide progress.
Short-term (1–3 years):
Enhance the coupling of heat and moisture models to better capture humidity dynamics and improve health-relevant indoor predictions.
Standardize benchmarking through consistent validation metrics, error reporting, and experimental calibration methods to improve comparability
Medium-term (3–5 years):
Expand model applications to diverse climates and housing types, particularly in underrepresented regions with extreme heat and humidity conditions.
Strengthen the integration of occupant behavior, HVAC operation, and envelope performance for more realistic residential simulations.
Long-term (5+ years):
Advance hybrid modeling that fuses physics-based and data-driven principles, such as gray-box and physics-informed neural networks, to balance interpretability and predictive power.
Develop open, interoperable frameworks and datasets that link physical models with sensor-based analytics for continuous validation and large-scale residential monitoring.
These priorities collectively define a roadmap for improving predictive accuracy, transferability, and real-world applicability, advancing the creation of energy-efficient, healthy, and climate-resilient homes.
Supplementary Material
Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.buildenv.2025.114125.
Funding & acknowledgments
Research reported in this publication was supported by the National Heart Lung and Blood Institute of the National Institutes of Health under award number 1R01HL164726–01.
Declaration of competing interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:
Dr Leah Schinasi reports financial support was provided by National Heart Lung and Blood Institute of the National Institutes of Health. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Data availability
No data was used for the research described in the article.
Abbreviation Full Term
- ANN
Artificial Neural Network
- AFN
Airflow Network
- ASHRAE
American Society of Heating, Refrigerating and Air-Conditioning Engineers
- BES
Building Energy Simulation
- BPS
Building Performance Simulation
- CFD
Computational Fluid Dynamics
- CNN
Convolutional Neural Network
- CNN-LSTM
Convolutional Neural Network – Long Short-Term Memory
- DEML
Deep Ensemble Machine Learning
- DL
Deep Learning
- DOE
U.S. Department of Energy
- GNN
Graph Neural Network
- HAMT
Heat and Moisture Transfer Module
- HVAC
Heating, Ventilation, and Air Conditioning
- IAQ
Indoor Air Quality
- IDA-ICE
IDA Indoor Climate and Energy
- IMG
Internal Moisture Generation
- LHS
Latin Hypercube Sampling
- LSTM
Long Short-Term Memory
- MAE
Mean Absolute Error
- MAPE
Mean Absolute Percentage Error
- ML
Machine Learning
- MOEA
Multi-Objective Evolutionary Algorithm
- MPC
Model Predictive Control
- MSE
Mean Squared Error
- NMBE
Normalized Mean Bias Error
- PCM
Phase Change Material
- PHPP
Passive House Planning Package
- PWARX
Piece-Wise Auto-Regressive with Exogenous Input
- R2
Coefficient of Determination
- RC Model
Resistance–Capacitance Model
- RECS
Residential Energy Consumption Survey
- RF
Random Forest
- RMSE
Root Mean Square Error
- RNN
Recurrent Neural Network
- SHGC
Solar Heat Gain Coefficient
- SRC
Standardized Regression Coefficients
- SVR / SVM
Support Vector Regression / Support Vector Machine
- TRNFLOW
TRNSYS Airflow Module
- TRNSYS
Transient System Simulation Tool
- WoS
Web of Science
- WUFI
Wärme Und Feuchte Instationär
Appendix
Appendix Table A:
Summary of Studies Adopting White Box Models.
| Study | Country | Building Type | Variable Predicted | Method | Quantitative Result | Key Contribution | Limitations |
|---|---|---|---|---|---|---|---|
| Gan et al. [102] | Singapore | Apartment building | T, RH, PMV, PPD | BIM + CFD + ML hybrid | PMV error 0.03–0.13; PPD error 0.35–5.52 % | Integrative BIM-CFD workflow for comfort | Sensitive to boundary settings and training data; comfort scope narrow. |
| Zhang et al. [103] | China | Bedroom (HVAC) | Indoor T | Semi-coupled CFD + lumped | Stable ~20 °C; faster convergence | Efficient transient HVAC simulation | Seasonal and geometric limits; no moisture coupling. |
| Winkler et al. [41] | USA | High-efficiency homes | Humidity (comfort indices) | EnergyPlus nodal comfort models | RH thresholds affect energy/comfort | Links RH thresholds to comfort/energy | Simplified comfort formulation; lacks physical airflow modeling. |
| Lee [104] | South Korea | Rural single-family | Indoor T, heating energy | EnergyPlus/DesignBuilder | GR room 4.2 °C cooler; >50 %o energy saved | Shows retrofit thermal/energy benefits | Limited sample; short seasonal period; simplified physics. |
| Tchawa et al. [105] | Cameroon | Prototype | Indoor T | Analytical (LOBATIN/HTSLM) | Modeled 14.9–30.2 °C vs EnergyPlus 20.6–24.9 °C | Provides lumped-parameter insight | Prototype scale; no moisture or airflow dynamics. |
| Lundqvist et al. [106] | Sweden | Multi-family | Indoor T, RH (PPD) | CFD (ANSYS CFX) + IDA ICE | Error <0.2 °C; PPD <10 % | Shows CFD-BPS comfort modeling | Single climate case; restricted validation. |
| Li and Zhao [42] | China | Rural farmhouse | Indoor T, RH | DesignBuilder + monitoring | r = −0.99 to −0.72 correlation | Demonstrates envelope buffering | Localized study; no airflow representation. |
| Murtyas et al. [36] | Malaysia | Terrace house | Indoor T | EnergyPlus | RMSE 0.4–1.03 °C | Shows passive cooling resilience | Fixed internal gains; uncertain climate data. |
| Martínez-Marino [74] | Germany | IEA Twin House | T, RH | TRNSYS/TRNFLOW + HAMFitPlus | RH RMSE 2.2–3.6 %; T RMSE 0.48–0.51 °C | Demonstrates calibrated hygrothermal simulation | Model accuracy depends on moisture-mass calibration. |
| Choi et al. [51] | Japan | Test room | T, RH | TRNSYS EMPD vs P-model | RH difference up to 38.8 % | Highlights advanced HAM coupling | Simplified coupling; limited validation. |
| Zhong et al. [65] | Japan | Machiya house | T, airflow | CFD RANS (SST k-ω) | Best match with SST k-ω model | Validates CFD for vernacular design | Coarse measurement detail; steady-flow assumption. |
| Leng et al. [64] | Malaysia | Terraced house | T, airflow | DesignBuilder CFD | CFD error ~15 %; T ↓ 2.06 °C | Demonstrates passive ventilation design | Simplified boundaries; location-specific findings. |
| Weerasuriya et al. [107] | Hong Kong | High-rise tower | T, airflow | eQuest + CFD + CONTAM | ~25 % energy savings | Shows hybrid NV strategy | Pressure-flow simplifications; no urban effects. |
| Garcia-Frometa et al. [49] | DR | Social housing | T, RH | DesignBuilder nodal | Wall effects vary by climate | Shows wall construction influence | Limited wall and ventilation variability. |
| Xiang et al. [81] | China | Apartment | Indoor T, MRT | EnergyPlus + field calibration | Cooling ↓ up to 1.1 °C; error ≤0.31 °C | Shows nocturnal ventilation benefits | Single dwelling; comfort validation absent. |
| Petrou et al. [46] | UK | Terraced/semidetached | Daily T | EnergyPlus + Bayesian | RMSE ↓ 2.5 → 0.6 °C | Provides scalable archetype calibration | Tool-dependent outcomes; limited calibration. |
| Sarna et al. [34] | Poland | Single-family | T, ACH | EnergyPlus + CONTAM | NMBE <1 %; CV(RMSE) ≤7 % | Rigorous calibration workflow | Simplified internal gains; qualitative moisture treatment. |
| Martin et al. [35] | USA | Single-family homes | RH, IMG | Lumped moisture balance | IMG RMSE = 2.6 kg d−1 | Provides baseline IMG rates | Occupant uncertainty; measurement error. |
| Yeretzian et al. [77] | Iran | Chamber sim | T, RH, VE | CFD + EnergyPlus | Cooling up to 6 °C; error ≤10 % | CFD + BPS validated for cooling design | Single archetype; limited variables analyzed. |
| Gori et al. [108] | UK | Historic houses | T, RH | WUFI, Delphin, EnergyPlus | Tools diverged in RH prediction | Shows tool-choice impact | Partial validation; restricted transferability. |
| Tian et al. [63] | China | Rural house | Indoor T, RH (qual.) | EnergyPlus | Mean T ↑ 2.2 °C | Demonstrates passive-solar integration | Trend-based calibration; envelope-only focus. |
| Bonello et al. [66] | Malta | Test chamber | Specific humidity, T | Transient CFD (Fluent) | Humidity error ±0.0015 g/g | Shows detailed CFD humidity modeling | Small test chamber; limited airflow |
| Kang et al. [48] | Australia | Passive House | Indoor T (overheating) | RC nodal (ISO 52,016) | CVRMSE <0.04; ASHRAE 14 compliant | Improves PHPP overheating prediction | accuracy. Quasi-steady assumption; simplified internal loads. |
| Petrou et al. [46] | UK | London flats | Indoor T | EnergyPlus vs IES VE | Tool-dependent compliance | Highlights tool sensitivity | Duplicate entry; same as above. |
| Heibati et al. [75] | Canada | Three-story house | RH | EnergyPlus + CONTAM + WUFI | Improved RH outcomes vs standalone | Demonstrates integrated coupling | Limited documentation of coupling accuracy. |
| Karyono et al. [109] | UK | Detached houses | T, RH, CO2 | Hybrid hygrothermal | T = 16–20 °C; RH rises with low heating | Captures hygrothermal across eras | Brief validation; incomplete detail. |
| Li et al. [47] | China | Passive solar house | Indoor T | Heat-balance ODE | Max deviation <0.6 ° C | Validates passive solar performance | Simplified passive model; no dynamic feedback. |
| Belloum et al. [72] | Algeria | Test cell | T, RH | EnergyPlus HAMT | RH error ±2.4 %; 6 h lag T | Demonstrates biobased envelope RH control | Passive cell only; no active system coupling. |
| Zhao et al. [56] | China | Prefab house | Indoor T | TRNSYS + MATLAB GA | Mean T higher from −1.8 → −0.2 °C | Provides retrofit insulation evidence | Divergence in extreme radiation; single-season data. |
| Egrna-Oller et al. [58] | Germany | Twin House | Indoor T | TRNSYS/TRNFLOW multizone | RMSE <1 °C; CV(RMSE) <4.8 % | Validates multizone models | Minor phase-lag prediction errors. |
| Zhang [80] et al. | China | Rural house (Shandong) | Indoor T | DesignBuilder | RMSE 0.15–0.29 °C; T ↑ 2.38 °C | Shows sunroom passive benefits | Single design variable; partial optimization scope. |
| Gamboa Loya et al. [50] | Mexico | Prototype dwelling with PCM walls | Indoor T, RH, energy | EnergyPlus CondFD + ML meta-model | Simulation-experiment deviation ≈5 %; ML energy error < 6 % | Shows PCM peak shaving and ML surrogate potential | Missing PCM properties; incomplete validation. |
| Gou et al. [110] | China | Passive solar house | Indoor temperature | DesignBuilder nodal | Low-e glazing gave most comfortable outcome | Demonstrates glazing system evaluation beyond U-value/SHGC | Limited time span; unclear measurement agreement. |
| Ju et al. [73] | China | Rural timber house | Comfort T, load shifting | TRNSYS PCM floor (Types 399/1270) | Room-T ± 0.8 °C; flexible pre-heating 5 h with ≤0.7 °C drift | Rare PCM-heat pump demand-response study | Comfort metric limited to temperature; simplified control. |
| Taing et al. [111] | Cambodia | Rural lightweight house | Operative T | DesignBuilder/EnergyPlus | Indoor T ↓ 2–4 °C under BDGC strategies | Validates design-guideline applicability in tropics | Prototype-based; lacks empirical validation. |
| Jiang et al. [112] | China | Steel prefab house | Indoor T swing; energy | EnergyPlus + simplified CFD roof | Capping roof raised mean T by 0.8 °C; reduced fluctuation | Adds retrofit roof options for severe-cold prefab | Partial data disclosure; ground simplifications. |
| Wang et al. [113] | China | Rural masonry house | Winter T, heating load | EnergyPlus parametric | OIS ↑ room T by 2.5–3.2 °C; radiant wall gave +1.2 °C | Introduces new passive-solar retrofit concept | No measured verification; seasonal restriction. |
| Martínez-Mariño et al. [74] | Spain | Lightweight test houses | Multi-zone T, RH | TRNSYS/TRNFLOW + calibration | RH error cut 69 %; RMSE ≈3 % RH, 0.5 °C | Demonstrates calibrated RH buffering | Dependent on sorption parameters; low-gain sensitivity. |
| Xu et al. [114] | China | Apartment (4-storey) | Indoor T; overheating; energy | EnergyPlus calibrated PCM model | CV(RMSE) ≤12 %; PCM ↓ T by 0.13–0.27 °C; overheating ↓ 9 % | Provides verified PCM retrofit benefit | Narrow benefit range; no moisture measurement. |
| Zhong et al. [115] | Hong Kong | High-rise flats | Indoor T, PM2.5 | EnergyPlus EMS + AFN | Reproduced open-duration within 10 %; ~70 Monte Carlo runs converged | Shows behavioural models’ effect on IAQ–thermal link | Static behavioral rules; limited occupant realism. |
| Cui et al. [116] | China | Rural brick house | Bedroom T; heating energy | DesignBuilder nodal | Retrofit ↑ mean T by 6.2 °C; error 𢑨1 °C | Demonstrates retrofit package benefits | Instrument error margin; short monitoring. |
| Tang et al. [55] | China | HPB archetypes | High-RH hours | EnergyPlus 4608 runs | >20 % RHH in small/medium HPBs; model bias <5 % | First large-scale health-oriented RH study | Uniform indoor assumptions; coarse moisture sourcing. |
| Grygierek et al. [57] | Poland | Two-storey house | Annual T, RH trend, CO2 | CONTAM + EnergyPlus + GA | Optimised CO2 <1000 ppm; NMBE 0.1 %, CVRMSE 2.4 % | Coupled IAQ–thermal GA workflow | Simplified natural ventilation; limited variability. |
| Sarmouk et al. [117] | Canada | Test cell with radiant slab | Room T; slab T | TRNSYS-Type56 + DoE | NMBE ≤5 %, CV(RMSE) ≤15 % | Links slab design to transient indoor climate | Simplified control logic; neglected solar coupling. |
| Athmani et al. [118] | Algeria | Masonry house | Indoor T; discomfort hours | TRNSYS calibrated | Cool-vent roof ↓ T by 4.95 °C; ↓ discomfort 45 % | Shows cool-roof passive mitigation | Single-month data; sparse sensors. |
| Liu [28] et al. [28] | China | Detached insulated house | Hourly T, RH | TRNSYS multi-zone calibrated | T = 20–24 °C; RH = 34–56 %; deviations ≤ ±1 °C/±2 % | Reliable dual-parameter prediction for cold-climate | Winter focus; high-altitude bias. |
| Yeretzian et al. [77] | Lebanon | 1960s masonry apartment | Indoor T, airflow; energy | EnergyPlus AFN + optimisation | HVAC use ↓ to 60 kWh/m2; deviations <10 % | Demonstrates mixed mode retrofit optimisation | Duplicate; same as above. |
| Gao et al. [71] | Hong Kong | Public rental block (40-storey) | Operative T; energy | EnergyPlus AFN zonal model | PCS yielded comfort/time gains; flats size sensitive | Nuanced PCS trade-offs in high-rise | PMV limits; comfort perception untested. |
| Santamaría et al. [54] | Spain | Façade prototypes | Indoor T under solar loads | Custom solver + test huts | Maintained indoor 5 °C cooler; error <5.5 % | First validation of water-flow glazing | Short-term test; maintenance effects omitted. |
| Vucicevic et al. [119] | Serbia | Post-war detached houses | Indoor T, wall T | TRNSYS multi-zone + PCM layer | Room air ↓2 °C; wall T ↓3.4 °C | Shows moderate PCM cooling | No experimental proof; coarse PCM modeling. |
| Angelotti et al. [120] | Italy | Apartment shell | Free-float indoor T | EnergyPlus, TRNSYS, IDA-ICE | Best auto-calibration RMSE = 0.3 °C | Compares calibration strategies | Small sample; limited occupant data. |
| Grygierek et al. [57] | Poland | Detached house | Indoor T; airflow proxy | EnergyPlus (10-zone) + fuzzy logic | 360–6140 h discomfort depending on NV | Demonstrates AI-aided NV strategy | Duplicate; simplified NV assumptions. |
| Firląg et al. [62] | Poland | Passive house | Operative T; loads | TRNSYS 18 zonal model | Overheated hours doubled (24 % → 44 %) | Shows climate-change sensitivity | No moisture or summer calibration. |
| Serageldin et al. [69] | Egypt | Two-storey concrete house | Indoor T; seasonal loads | CFD (Fluent) + TRNSYS | Summer T ↓5–9 °C; load = 18.8 kWh/m2 | Hybrid CFD-BES for passive systems | One dwelling; limited variable range. |
| Tong et al. [37] | Singapore | High-rise apartment | Indoor T; cooling load | DesignBuilder calibrated | Peak T ↓1.8 °C; cooling ↓7 % | Demonstrates faąade-driven cooling | Constant ventilation rate; interaction effects missing. |
| Kadri et al. [59] | Algeria | Adobe dwelling | Operative T; cooling energy | TRNSYS calibrated | T ↓5 °C; cooling ↓66 % | Shows double-skin reflective roof | Material aging and climate transfer untested. |
| Rincón et al. [52] | Spain | Earthbag house | Indoor T, RH, U-value | EnergyPlus + field | Temp swing ↓90 %; MAE = 1.15 °C | Provides evidence for low-tech envelope | Partial modeling chain; climatic constraint. |
| Zhang et al. [121] | China | Vernacular house | Summer T, RH | DesignBuilder parametric + field | RMSE 1.5–2.0 °C; MAPE 9–11 % | Provides rare humid mountain dataset | Hot-season focus; limited generalization. |
| Radujković et al. [60] | Belgium | Apartment façade | Operative T; heating energy | DesignBuilder + EnergyPlus | T i3.5 °C summer; T ↑1.4 °C winter | Demonstrates vegetation façade retrofit | Few sensors; indoor-surface data absent. |
| Abbaas et al. [61] | Jordan | Apartment blocks | Indoor T, MRT, RH | EnergyPlus/OpenStudio | Night NV ↓ indoor T by 4–10 °C | Shows shape-sensitive NV outcomes | Short validation period; simplified climate input. |
| Vila-Hernoandez et al. [33] | Mexico | Social-housing module | Operative T; roof heat flux | EnergyPlus calibrated | Roof ↓ T by up to 4.7 °C; cooling ↓99 % | Expands Latin-American roof data | Simplified roof substrate physics; single-room test. |
| Connolly et al. [122] | USA | Multi-family apartment | Indoor T↑2.1 °C; cooling energy | EnergyPlus nodal | Indoor T ↑2.1 °C; cooling ↑34–50 % | Quantifies future climate-demand impacts | One archetype; single-model climate scenario. |
| Vadiee et al. [43] | Sweden | 6-storey apartment | Operative T; heating demand | TRNSYS, EnergyPlus, IDA-ICE, VIP | Heating demand 38–44 kWh/m2 | Shows inter-tool differences | Tool-setup differences; unverified humidity behavior. |
| Radujkovioć et al. [60] | Belgium | Residential building | Indoor temperature | EnergyPlus via DesignBuilder | T ↓3.5 °C (summer); T ↑1.4 °C (winter) | Reinforces passive wall retrofit evidence | Duplicate; same as above. |
| Van Hove et al. [76] | Netherlands | Zero-energy house | Indoor T, RH, CO2 | Dymola/Modelica multi-zone | RMSE T < 1 °C; RH error <5 % | High-resolution dynamic validation | Sensor failure; long runtime; occupancy dependence. |
| Nyame-Tawiah et al. [123]. | Ghana | Test cells | Indoor temperature | EnergyPlus + field | Indoor T ↓1.1 ° C; validated on 9 cells | Provides tropical green-roof data | No ventilation coupling; minimal vegetation diversity. |
| Ayoobi et al. [70] | Afghanistan | Test room | Indoor T (NV) | EnergyPlus + CFD hybrid | Windcatcher NV reduced cooling load | Hybrid method for arid NV | Weak calibration; electrical data gaps. |
| Bonello et al. [67] | Malta | Test chamber | RH | Transient CFD | RH validated ±0.001 g/g; 31 % RH variation | Demonstrates spatial RH differences | Small-scale chamber; limited spatial realism. |
| Belpoliti et al. [68] | UAE | Demonstration house | Indoor temperature | CFD + EnergyPlus calibrated | NV ↓ T by ~0.7 °C | Shows CFD-based passive cooling | Short monitoring; external flow not modeled. |
Appendix Table B.
Summary of Studies Adopting Black Box Models.
| Study | Country | Building Type | Variable Predicted | Method | Quantaitative Result | Key Findings | Limitations |
|---|---|---|---|---|---|---|---|
| Weng et al. [83] | UK | Two-storey terraced house | Operative temperature (+ humidity implied) | LSTM-RNN with ventilation scenarios | 70 % < 1 °C error; R2 = 0.956 | Anticipates natural ventilation; ≥ 20 % RMSE reduction vs baselines | Limited to one house; no direct humidity modeling |
| Song et al. [86] | China | Mid-rise apartments (district heating) | Indoor temperature | HAGRU deep neural network with attention | Accuracy ≈ 98.4 % | Two-layer attention mechanism improved interpretability; robust 4 h forecasts | District-heating only; not validated on cooling |
| Sözer et al. [40] | Turkey | Elderly home (8-storey, 18,108 m2) | Temperature + humidity | ANN (3-layer MLP, multi-zone) | Error ≤ 1.3 °C; ≤ 2 % RH | Reduced monitoring by 77 %; validated vs ASHRAE 14; guided simulation calibration | Dataset limited to 1 heating season; one site only |
| Potočnik et al. [85] | Slovenia | Simulated single-family dwelling | Indoor temperature | ARX, ANN, ELM | RMSE ≈ 0.065 ° C (NN-BR) | Non-linear ML outperformed linear ARX; future weather improved results | Simulated dwelling only; no RH prediction |
| Benzaama et al. [90] | France | Student dormitory (1990s) | Room-air temperature | Piecewise affine ARX (PWARX) | 78.5 % accuracy | Lightweight operational-data model; captured 70 % heating, 20 % solar effects | Simplified piecewise structure; no cooling/humidity consideration |
| Chen et al. [84] | China | Small offices (naturally ventilated) | Temp + RH | CNN-LSTM transfer learning | MSE = 0.16 °C; 2.52 % RH | Reduced commissioning time; robust 6 h forecasts | Limited to two office sites; generalization untested |
| Al-Obeidat et al. [89] | USA | Two single-family houses | Indoor temperature | Ridge & Lasso regression | MAE 0.12 °C; RMSE 0.18 °C (3 h); <1.8 °C (48 h) | Demonstrated regression accuracy; adequate for smart-home forecasting | Focused only on temperature; excludes RH and energy impact |
| Ramadaon et al. [82] | France | Laboratory test cell | Indoor temperature | ANN, RF, XGBoost, etc. | ANN RMSE = 0.081 °C; R2 = 0.9997 | ML methods surpass RC; sub-0.2 °C RMSE over 60 min | Artificial test cell only; not applied in real homes |
| Alhamayani et al. [31] | USA | University-owned residences | Temp + comfort (energy savings) | LSTM + PMV comfort model | 33–47 % savings with solar input | Framework for comfort-based predictive control | Assumes PMV as universal metric; only cooling season |
| Delcroix et al. [39] | Canada | 7 electrically heated houses | Temp + load | Coupled ANNs | RMSE 0.29–1.09 °C | Multi-house dataset; strong benchmark for heating load predictions | Seasonal scope only; no humidity inclusion |
| Huchuk et al. [30] | USA/ Canada | 1000 single-family homes | Indoor temperature | Ridge, Lasso, ARX, RF | Ridge lowest errors; climate differences noted | Showed large-scale thermostat forecasting validity | Climate variability increased errors; temperature only |
| Liu et al. [124] | Norway | Single-family house (100 m2) | Humidity efficiency | Gaussian Process Regression | Peak eff. 68 %; avg 19 % | Surrogate humidity model; high accuracy | Envelope-specific; limited generalization |
| Kim et al. [99] | Switzerland | Multifamily apartments | Temp + humidity | ANN (FF + RNN, LM-BP) | Error ≈ 3.36–6.12 % | Good IAQ prediction; validated against measurements | Narrow case study; seasonal limits |
| Yu et al. [92] | Australia | 25 mixed-use (residential subset) | Indoor temperature | Deep Ensemble ML | RMSE 0.125–0.886 °C; R2 ≈ 0.96 | Stable across climates/seasons; robust generalization | Data from sensors only; no humidity prediction |
| Boubouh et al. [87] | USA | Single-family homes (1000) | Indoor temperature | RNN/LSTM | High accuracy; 384 × less energy | Efficient personalized training approach | Tested in US only; generalizability open |
| Espinosa et al. [32] | Spain | Domotic house (residential) | Temp + NO2 | MOEA (RF + LSTM surrogates) | Error ↓ ≈ 13 % | Improved short-horizon forecasts; efficient runtime | Limited to one site; computationally intensive |
| Laukkarinen et al. [44] | Finland | Six single-family houses | Indoor temperature | Ensemble ML (RF, XGBoost, etc.) | MAE 0.5–0.8 °C | Benchmarked ensemble vs SVR; trees most robust | Campaign short-term only; no RH data |
| Shrestha et al. [125] | Japan | Traditional samurai house | Temperature | SVR + RF + feature importance | Identified inter-room influences | Non-destructive rapid assessment for retrofits | 76 h data only; no humidity record |
| Pergantis et al. [91] | USA | Detached single-family | Indoor humidity | Latent regression (SHR model) | Cut peak power vs constant SHR | Showed RH modeling improves demand response | Limited RH benefit in some cases |
| Hu et al. [97] | Japan | Residential AC lab | Temp + RH + energy | Dynamic GNN + GRU + self-attention | Avg R2 = 0.935; +25 % vs LSTM | Multi-task robust prediction across AC performance variables | High complexity; computational demand |
| Alshammari [96] | Saudi Arabia | Smart home (250 m2, HVAC) | Temperature | ANN (MLP, tuned) | MAE 0.056 °C; RMSE 0.071 °C; R2 = 0.997 | Lightweight ANN for embedded smart-home platforms | Limited to one house; no RH modeled |
| Qi et al. [88] | China | Central heating apartments | Temperature | CONST constraint-guided NN | Interpretability × 8 higher | Physically plausible forecasts with better interpretability | Heating-only validation; no RH prediction |
| Yuksek et al. [100] | Turkiye | Test house (48 m2) | Temperature | ANN, GPR, SVM, ANFIS | Best GPR R2 = 0.998; RMSE = 0.134 °C | Benchmarked multiple ML methods; GPR, ANN most accurate | Very small test house; limited validation |
| Jindal et al. [29] | China | Community blocks (district heating) | Temperature | Deep RL (MTDN + SAC) | Stable control validated vs history | Innovative RL + DL for district heating | District-heating only; cooling ignored |
| Shakhovska et al. [98] | Ukraine | Smart infrastructure | Indoor temperature | LSTM + boosting | Test losses 0.0005–0.028 | Robust workflow for time-series forecasting | Building details unclear |
| Farahani et al. [38] | Finland | Multifamily apartments | Indoor temperature | XGBoost, LSTM, MLR | MAE 0.23 °C; MSE 0.12 | Ensembles better overall; LSTM better in high T | Focused on summer season only |
| Sung et al. [94] | Taiwan | Smart home with sensors | Temp + RH + CO2 | CNN-LSTM + KNN imputation | Temp RMSE = 0.042 °C; RH err. = 0.0415 % | High-frequency multi-variable forecasting; robust IAQ predictions | Limited to one IoT system |
Footnotes
Note: The term “humidity only” in refers to studies where humidity-related variables (e.g., relative humidity, specific humidity, or humidity ratio) were the primary outputs analyzed, while air temperature was still internally solved within the coupled heat- and mass-balance equations. These therefore represent humidity-focused predictions rather than independent humidity modeling.
Note: The term “humidity only” in refers to studies where humidity-related variables (e.g., relative humidity, specific humidity, or humidity ratio) were the primary outputs analyzed, while air temperature was still internally solved within the coupled heat- and mass-balance equations. These therefore represent humidity-focused predictions rather than independent humidity modeling.
CRediT authorship contribution statement
Chima Cyril Hampo: Writing – review & editing, Writing – original draft, Investigation, Formal analysis, Data curation, Conceptualization. Leah H. Schinasi: Writing – review & editing, Supervision, Resources, Funding acquisition. Simi Hoque: Writing – review & editing, Supervision, Resources, Project administration.
Declaration of generative AI and AI-assisted technologies in the writing process
During the preparation of this work the authors used ChatGPT to improve the readability and organization of the manuscript. After using this tool/service, the authors reviewed and edited the content as needed and takes full responsibility for the content of the publication.
References
- [1].Seneviratne SI, et al. , Weather and climate extreme events in a changing climate. 2021. [Google Scholar]
- [2].Wu J, et al. , Estimation and uncertainty analysis of impacts of future heat waves on mortality in the eastern United States, Environ. Health Perspect 122 (1) (2014) 10–16. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [3].Brown EG Jr, Rodriquez M, and Chapman R, Preparing California for extreme heat. 2013. [Google Scholar]
- [4].Stafoggia M, et al. , Vulnerability to heat-related mortality: a multicity, population-based, case-crossover analysis, Epidemiology. (Fairfax) (2006) 315–323. [Google Scholar]
- [5].Wheeler K, et al. , Heat illness and deaths—New York City, 2000–2011, Morbidity and Mortality Weekly Report 62 (31) (2013) 617. [PMC free article] [PubMed] [Google Scholar]
- [6].Fouillet A, et al. , Excess mortality related to the August 2003 heat wave in France, Int. Arch. Occup. Environ. Health 80 (2006) 16–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [7].Din A and Brotas L, The evaluation of the variables of domestic overheating in the UK under TM52 using a future climate model-guidance for designers. 2016. [Google Scholar]
- [8].Guía A, Guías de CIBSE. [Google Scholar]
- [9].Gronlund C, Racial and socioeconomic disparities in heat-related health effects and their mechanisms: a review, Curr. Epidemiol. Rep 1 (3) (2014) 165–173. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [10].Benmarhnia T, et al. , Vulnerability to heat-related mortality, Epidemiology. (Fairfax) 26 (6) (2015) 781–793. [Google Scholar]
- [11].Hampo CC, Schinasi LH, Hoque S, Surviving Indoor Heat Stress in United States: A comprehensive Review Exploring the Impact of Overheating On the Thermal comfort, health, and Social Economic Factors of Occupants, Heliyon, 2024. [Google Scholar]
- [12].Pathan A, et al. , Monitoring summer indoor overheating in the London housing stock, Energy Build. 141 (2017) 361–378. [Google Scholar]
- [13].Zhivov A, et al. , Requirements for building thermal conditions under normal and emergency operations in extreme climates, ASHRAe J. 127 (1) (2021) 693. [PMC free article] [PubMed] [Google Scholar]
- [14].Gupta V, Sarkar A, Jana A, Impact of occupant behavior on indoor thermal comfort and ventilation patterns in social housing of Mumbai, India: observation from experiments and household surveys, in: Proceedings of the SimBuild Conference, 2024. [Google Scholar]
- [15].Foucquier A, et al. , State of the art in building modelling and energy performances prediction: a review, Renewable and Sustainable Energy Reviews 23 (2013) 272–288. [Google Scholar]
- [16].Zhao H.-x, Magoulès F, A review on the prediction of building energy consumption, Renewable and Sustainable Energy Reviews 16 (6) (2012) 3586–3592. [Google Scholar]
- [17].Yu J, Chang W-S, Dong Y, Building energy prediction models and related uncertainties: a review, Buildings 12 (8) (2022) 1284. [Google Scholar]
- [18].Chen Y, et al. , Physical energy and data-driven models in building energy prediction: a review, Energy Reports 8 (2022) 2656–2671. [Google Scholar]
- [19].Hamdaoui M-A, et al. , A review on physical and data-driven modeling of buildings hygrothermal behavior: models, approaches and simulation tools, Energy Build. 251 (2021) 111343. [Google Scholar]
- [20].Al Mindeel T, Spentzou E, Eftekhari M, Energy, thermal comfort, and indoor air quality: multi-objective optimization review, Renewable and Sustainable Energy Reviews 202 (2024) 114682. [Google Scholar]
- [21].Lu C, Li S, Lu Z, Building energy prediction using artificial neural networks: a literature survey, Energy Build. 262 (2022) 111718. [Google Scholar]
- [22].Ngarambe J, Yun GY, Santamouris M, The use of artificial intelligence (AI) methods in the prediction of thermal comfort in buildings: energy implications of AI-based thermal comfort controls, Energy Build. 211 (2020) 109807. [Google Scholar]
- [23].Feng Y, et al. , Data-driven personal thermal comfort prediction: a literature review, Renewable and Sustainable Energy Reviews 161 (2022) 112357. [Google Scholar]
- [24].Saini J, Dutta M, Marques G, Indoor air quality prediction systems for smart environments: a systematic review, J. Ambient. Intell. Smart. Environ 12 (5) (2020) 433–453. [Google Scholar]
- [25].Swan LG, Ugursal VI, Modeling of end-use energy consumption in the residential sector: a review of modeling techniques, Renewable and sustainable energy reviews 13 (8) (2009) 1819–1835. [Google Scholar]
- [26].ASHRAE, A., Standard 55-thermal environmental conditions for human occupancy. 2017. [Google Scholar]
- [27].Liu S, et al. , Personal thermal comfort models with wearable sensors, Build. Environ 162 (2019) 106281. [Google Scholar]
- [28].Liu Z, et al. , Performance and feasibility study of low energy building for pleasant indoor comfort: a case study on Qinghai-Tibetan Plateau, Sustain. Cities. Soc 50 (2019) 101691. [Google Scholar]
- [29].Jindal L, et al. , Deep learning-based heat optimization techniques for forecasting indoor temperature changes, Spatial Information Research 32 (1) (2024) 107–117. [Google Scholar]
- [30].Huchuk B, Sanner S, O’Brien W, Evaluation of data-driven thermal models for multi-hour predictions using residential smart thermostat data, J. Build. Perform. Simul 15 (4) (2022) 445–464. [Google Scholar]
- [31].Alhamayani AD, Sun Q, Hallinan KP, Estimating smart wi-fi thermostat-enabled thermal comfort control savings for any residence, Clean Technologies 3 (4) (2021) 743–760. [Google Scholar]
- [32].Espinosa R, Jiménez F, Palma J, Multi-surrogate assisted multi-objective evolutionary algorithms for feature selection in regression and classification problems with time series data, Inf Sci (Ny) 622 (2023) 1064–1091. [Google Scholar]
- [33].Avila-Hernández A, et al. , Test box experiment and simulations of a green-roof: thermal and energy performance of a residential building standard for Mexico, Energy Build. 209 (2020) 109709. [Google Scholar]
- [34].Sarna I, Ferdyn-Grygierek J, Grygierek K, Thermal model validation process for building environment simulation: a case study for single-family house, Atmosphere (Basel) 13 (8) (2022) 1295. [Google Scholar]
- [35].Martin E, et al. , Estimating internal moisture generation rates in recently constructed, occupied homes in the southeastern United States, Sci. Technol. Built. Environ 30 (7) (2024) 833–857. [Google Scholar]
- [36].Murtyas S, et al. , Thermal comfort in a two-storey malaysian terrace house: are passive cooling methods sufficient in present and future climates? Journal of Building Engineering 96 (2024) 110412. [Google Scholar]
- [37].Tong S, et al. , Impact of façade design on indoor air temperatures and cooling loads in residential buildings in the tropical climate, Energy Build. 243 (2021) 110972. [Google Scholar]
- [38].Farahani AV, et al. , Predicting summer indoor temperatures in Nordic apartments considering heatwaves forecasts, Energy Build. 336 (2025) 115630. [Google Scholar]
- [39].Delcroix B, et al. , Coupling of neural models for predicting indoor temperatures and heating loads in buildings. Building Simulation 2021, IBPSA, 2021. [Google Scholar]
- [40].Sözer H, Aldin SS, Predicting the indoor thermal data for heating season based on short-term measurements to calibrate the simulation set-points, Energy Build. 202 (2019) 109422. [Google Scholar]
- [41].Winkler J, Munk J, Woods J, Sensitivity of occupant comfort models to humidity and their effect on cooling energy use, Build. Environ 162 (2019) 106240. [Google Scholar]
- [42].Li K, Zhao T, The effect of envelope components on thermal performance of rural houses in Hubei, China, Indoor and built environment 28 (9) (2019) 1272–1287. [Google Scholar]
- [43].Vadiee A, Dodoo A, Gustavsson L, A comparison between four dynamic energy modeling tools for simulation of space heating demand of buildings. Cold Climate HVAC Conference, Springer, 2018. [Google Scholar]
- [44].Laukkarinen A, Vinha J, Long-term prediction of hourly indoor air temperature using machine learning, Energy Build. 325 (2024) 114972. [Google Scholar]
- [45].Petrou G, et al. , Can the choice of building performance simulation tool significantly alter the level of predicted indoor overheating risk in London flats? Building Services Engineering Research and Technology 40 (1) (2019) 30–46. [Google Scholar]
- [46].Petrou G, et al. , Development of a Bayesian calibration framework for archetype-based housing stock models of summer indoor temperature, J. Build. Perform. Simul (2024) 1–20. [Google Scholar]
- [47].Li J, Zhang Y, Yue T, A new approach for indoor environment design of passive solar buildings in plateau areas, Sustainable Energy Technologies and Assessments 63 (2024) 103669. [Google Scholar]
- [48].Kang Y, et al. , Optimizing the Passive House Planning Package simulation tool: a bottom-up dynamic approach to reduce building performance gap, Energy Build. 276 (2022) 112512. [Google Scholar]
- [49].García Frómeta Y, Ruiz Valero L, Cuadrado Rojo J, Indoor temperature and relative humidity assessment of three construction systems for Dominican social housing in different micro-climates: a modelling study, Indoor and Built Environment 28 (5) (2019) 693–710. [Google Scholar]
- [50].Gamboa Loya B, et al. , Assessment of thermal comfort with phase change materials in a standard house for different Mexican climates: a simulation study using EnergyPlus, Revista Mexicana de Ingeniería Química (2024) 23. [Google Scholar]
- [51].Choi Y, et al. , Influence of the moisture driving force of moisture adsorption and desorption on indoor hygrothermal environment and building thermal load, Energy Build. 253 (2021) 111501. [Google Scholar]
- [52].Rincón L, et al. , Analysis of the thermal behavior of an Earthbag Building in Mediterranean continental climate: monitoring and simulation, Energies. (Basel) 13 (1) (2019) 162. [Google Scholar]
- [53].Xu Y, et al. , Integrating phase-change materials to reduce overheating risk in residential buildings in cold regions of China, Constr. Build. Mater 475 (2025) 141153. [Google Scholar]
- [54].Santamaria B, et al. , Experimental validation of water flow glazing: transient response in real test rooms, sustainability (Switzerland). 12 (2020) 1–23. [Google Scholar]
- [55].Tang S, et al. , Unhealthy indoor humidity levels associated with ventilation rate regulations for high-performance buildings in China, Build. Environ 177 (2020) 106839. [Google Scholar]
- [56].Zhao J, et al. , Thermal performance and passive energy saving optimization of prefabricated houses in Xinjiang region, Case Studies in Thermal Engineering 70 (2025) 106082. [Google Scholar]
- [57].Grygierek K, Sarna I, Impact of passive cooling on thermal comfort in a single-family building for current and future climate conditions, Energies. (Basel) 13 (20) (2020) 5332. [Google Scholar]
- [58].Eguía-Oller P, et al. , Empirical validation of a multizone building model coupled with an air flow network under complex realistic situations, Energy Build. 249 (2021) 111197. [Google Scholar]
- [59].Kadri M, Bouchair A, Laafer A, The contribution of double skin roof coupled with thermo reflective paint to improve thermal and energy performance for the ‘Mozabit’houses: case of Beni Isguen’s Ksar in southern Algeria, Energy Build. 256 (2022) 111746. [Google Scholar]
- [60].Radujković M, Versele A, Breesch H, Exploratory analysis of a novel modular green wall’s impact on indoor temperature and energy consumption in residential buildings: a case study from Belgium, Energies. (Basel) 17 (21) (2024) 5267. [Google Scholar]
- [61].Abbaas ES, et al. , Effect of natural ventilation on thermal performance of different residential building forms in the hot-dry climate of Jordan, Pertanika J Sci Technol (1) (2024) 32. [Google Scholar]
- [62].Firląg S, Miszczuk A, Witkowski B, Analysis of climate change and its potential influence on energy performance of building and indoor temperatures, part 1: climate change scenarios, Archives of Civil Engineering 67 (3) (2021). [Google Scholar]
- [63].Tian H, et al. , Energy saving retrofit of rural house based on the joint utilization of solar collector and attached sunspace, Energy Build. 299 (2023) 113591. [Google Scholar]
- [64].Leng PC, et al. , The impact of air well geometry in a Malaysian single storey terraced house, Sustainability. 11 (20) (2019) 5730. [Google Scholar]
- [65].Zhong W, Zhang T, Tamura T, Cfd simulation of convective heat transfer on vernacular sustainable architecture: validation and application of methodology, Sustainability. 11 (15) (2019) 4231. [Google Scholar]
- [66].Bonello M, Micallef D, Borg SP, Humidity micro-climate characterisation in indoor environments: a benchmark study, Journal of Building Engineering 28 (2020) 101013. [Google Scholar]
- [67].Bonello M, Micallef D, Borg SP, Humidity distribution in high-occupancy indoor micro-climates, Energies. (Basel) 14 (3) (2021) 681. [Google Scholar]
- [68].Belpoliti V, et al. , Assessment of natural ventilation techniques by means of measurements and retrospective CFD simulation on a test building, Journal of Architectural Engineering 30 (2) (2024) 04024011. [Google Scholar]
- [69].Serageldin AA, et al. , Solar chimney combined with earth to-air heat exchanger for passive cooling of residential buildings in hot areas, Solar Energy 206 (2020) 145–162. [Google Scholar]
- [70].Ayoobi AW, Ekimci BG, Inceoğlu M, A comparative study of sustainable cooling approaches: evaluating the performance of natural ventilation strategies in arid and semi-arid regions, Buildings 14 (12) (2024) 3995. [Google Scholar]
- [71].Gao K, et al. , Balancing thermal comfort and energy efficiency in high-rise public housing in Hong Kong: insights and recommendations, J. Clean. Prod 437 (2024) 140741. [Google Scholar]
- [72].Belloum R, et al. , Hygrothermal performance assessment of a bio-based building made with date palm concrete walls, Build. Environ 223 (2022) 109467. [Google Scholar]
- [73].Ju H, et al. , Heating performance of PCM radiant floor coupled with horizontal ground source heat pump for single-family house in cold zones, Renew. Energy 235 (2024) 121306. [Google Scholar]
- [74].Martínez-Mariño S, et al. , Simulation and validation of indoor temperatures and relative humidity in multi-zone buildings under occupancy conditions using multi-objective calibration, Build. Environ 200 (2021) 107973. [Google Scholar]
- [75].Heibati S, Maref W, Saber HH, Assessing the energy, indoor air quality, and moisture performance for a three-story building using an integrated model, part three: development of integrated model and applications, Energies. (Basel) 14 (18) (2021) 5648. [Google Scholar]
- [76].Matthias Van Hove M, ME JBB, Indoor climate prediction performance of dynamic BES-models in Dymola. ASHRAE Topical Conference Proceedings, American Society of Heating, Refrigeration and Air Conditioning Engineers, Inc, 2021. [Google Scholar]
- [77].Yeretzian A, Karam J, Optimizing the use of natural ventilation while reducing energy cost_case study for retrofitting an old residential apartment in a Mediterranean city, Energy Build. 296 (2023) 113371. [Google Scholar]
- [78].Qin M, Yang J, Evaluation of different thermal models in EnergyPlus for calculating moisture effects on building energy consumption in different climate conditions. Building Simulation, Springer, 2016. [Google Scholar]
- [79].Rode C, Grau K, Moisture buffering and its consequence in whole building hygrothermal modeling, J. Build. Phys 31 (4) (2008) 333–360. [Google Scholar]
- [80].Zhang S, Jiang J, Gao W, Application evaluation and optimization of the sunroom in rural residential houses in Southeast Shandong Province, China, Solar Energy 251 (2023) 208–222. [Google Scholar]
- [81].Xiang H, Li J, Impact of night ventilation on indoor thermal environment of residential buildings under the dual carbon target: a case study of Xi’an, Buildings 14 (8) (2024) 2459. [Google Scholar]
- [82].Ramadan L, et al. , Use of machine learning methods for indoor temperature forecasting, Future Internet. 13 (10) (2021) 242. [Google Scholar]
- [83].Weng K and Mourshed M, RNN-based forecasting of indoor temperature in a naturally ventilated residential building. 2019. [Google Scholar]
- [84].Chen Y, et al. , Transfer learning with deep neural networks for model predictive control of HVAC and natural ventilation in smart buildings, J. Clean. Prod 254 (2020) 119866. [Google Scholar]
- [85].Potočnik P, et al. , Neural network, ARX, and extreme learning machine models for the short-term prediction of temperature in buildings. Building Simulation, Springer, 2019. [Google Scholar]
- [86].Song J, et al. , An indoor temperature prediction framework based on hierarchical attention gated recurrent unit model for energy efficient buildings, IEEe Access. 7 (2019) 157268–157283. [Google Scholar]
- [87].Boubouh K, et al. , Efficacy of temporal and spatial abstraction for training accurate machine learning models: a case study in smart thermostats, Energy Build. 296 (2023) 113377. [Google Scholar]
- [88].Qi D, et al. , Spatio-temporal consistency enhanced differential network for interpretable indoor temperature prediction, in: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024. [Google Scholar]
- [89].Al-Obeidat F, Spencer B, Alfandi O, Consistently accurate forecasts of temperature within buildings from sensor data using ridge and lasso regression, Future Generation Computer Systems 110 (2020) 382–392. [Google Scholar]
- [90].Hichem BENZAAMA M, et al. , Data-driven approach for modeling the thermal dynamics of residential buildings using a PieceWise ARX model. Building Simulation 2019, IBPSA, 2019. [Google Scholar]
- [91].Pergantis EN, et al. , Humidity-aware model predictive control for residential air conditioning: a field study, Build. Environ 266 (2024) 112093. [Google Scholar]
- [92].Yu W, et al. , Sensor-based indoor air temperature prediction using deep ensemble machine learning: an Australian urban environment case study, Urban. Clim 51 (2023) 101599. [Google Scholar]
- [93].Jogunola O, et al. , CBLSTM-AE: a hybrid deep learning framework for predicting energy consumption, Energies. (Basel) 15 (3) (2022) 810. [Google Scholar]
- [94].Sung G-M, et al. , An artificial intelligence home monitoring system that uses CNN and LSTM and is based on the Android studio development platform, Applied Sciences (2076-3417) (3) (2025) 15. [Google Scholar]
- [95].Yang S, et al. , An adaptive robust model predictive control for indoor climate optimization and uncertainties handling in buildings, Build. Environ 163 (2019) 106326. [Google Scholar]
- [96].Alshammari T, Using artificial neural networks with GridSearchCV for predicting indoor temperature in a smart home, Engineering, Technology & Applied Science Research 14 (2) (2024) 13437–13443. [Google Scholar]
- [97].Hu Z, et al. , Self-learning dynamic graph neural network with self-attention based on historical data and future data for multi-task multivariate residential air conditioning forecasting, Appl. Energy 364 (2024) 123156. [Google Scholar]
- [98].Shakhovska N, et al. , Innovative machine learning approaches for indoor air temperature forecasting in smart infrastructure, Sci. Rep 15 (1) (2025) 47. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [99].Kim MK, et al. , Prediction and correlation analysis of ventilation performance in a residential building using artificial neural network models based on data-driven analysis, Sustain. Cities. Soc 83 (2022) 103981. [Google Scholar]
- [100].Yüksek AG, et al. , Modeling the indoor temperature depending on insulation thickness using machine learning methods, The European Physical Journal Plus 139 (10) (2024) 887. [Google Scholar]
- [101].Baldwin JW, et al. , Humidity’s role in heat-related health outcomes: a heated debate, Environ. Health Perspect 131 (5) (2023) 055001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [102].Gan VJ, et al. , BIM and data-driven predictive analysis of optimum thermal comfort for indoor environment, Sensors 21 (13) (2021) 4401. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [103].Zhang L, et al. , Study of transient indoor temperature for a HVAC room using a modified CFD method, Energy Procedia 160 (2019) 420–427. [Google Scholar]
- [104].Lee K-I, Improvement of indoor thermal environments through green refurbishment, Sustainability. 12 (12) (2020) 4933. [Google Scholar]
- [105].Tchawa LPN, et al. , Heat transfer software for local materials in Cameroon (HTSLM 1.0): simplified building energy software based on local realities, Discover Energy 5 (1) (2025) 4. [Google Scholar]
- [106].Lundqvist P, Risberg M, Westerlund L, Indoor thermal climate after energy efficiency measures of a residential building in a sub-Arctic region: comparing ANSYS CFX and IDA ICE methods, Indoor and Built Environment 31 (3) (2022) 732–744. [Google Scholar]
- [107].Weerasuriya AU, et al. , A holistic framework to utilize natural ventilation to optimize energy performance of residential high-rise buildings, Build. Environ 153 (2019) 218–232. [Google Scholar]
- [108].Gori V, et al. , Assessing the role of simulation tool selection for the evaluation of heat and moisture balance in historic buildings. IOP Conference Series: Earth and Environmental Science, IOP Publishing, 2021. [Google Scholar]
- [109].Karyono K, et al. , The role of hygrothermal modelling for different housing typologies by estimating indoor relative humidity, energy usage and anticipation of fuel poverty, Build. Environ 207 (2022) 108468. [Google Scholar]
- [110].Gou W, Su H, Analysis of influence of glazing systems on indoor environment of a passive solar building in Lhasa. IOP Conference Series: Earth and Environmental Science, IOP Publishing, 2020. [Google Scholar]
- [111].Taing K, et al. , Bioclimatic design guidelines for design decision support to enhance residential building thermal performance in tropical regions, Sustainability (Switzerland) (4) (2025) 17. [Google Scholar]
- [112].Jiang W, et al. , Energy-saving retrofits of prefabricated house roof in severe cold area, Energy 254 (2022) 124455. [Google Scholar]
- [113].Wang W, et al. , Numerical investigation on the impact of an on-top sunspace passive heating approach for typical rural buildings in northern China, Solar Energy 186 (2019) 300–310. [Google Scholar]
- [114].Xu Y, et al. , Energy-saving and overheating-mitigation effects of phase change materials on buildings in cold regions, Energy Build. (2025) 116284. [Google Scholar]
- [115].Zhong X, Ridley IA, Verification of behavioural models of window opening: the accuracy of window-use pattern, indoor temperature and indoor PM2. 5 concentration prediction. Building Simulation, Springer, 2020. [Google Scholar]
- [116].Cui Y, et al. , Indoor temperature improvement and energy-saving renovations in rural houses of China’s cold region—A case study of Shandong Province, Energies. (Basel) 13 (4) (2020) 870. [Google Scholar]
- [117].Sarmouk MDE, et al. , On the mitigation of the overheating caused by floor heating systems exposed to long direct solar radiation and high occupation rates, Solar Energy 274 (2024) 112565. [Google Scholar]
- [118].Athmani W, et al. , The potential of using passive cooling roof techniques to improve thermal performance and energy efficiency of residential buildings in hot arid regions, Buildings 13 (1) (2023) 21. [Google Scholar]
- [119].Vučićević BS, et al. , Indoor temperatures affected by phase-change materials, Thermal Science 27 (6 Part A) (2023) 4513–4523. [Google Scholar]
- [120].Angelotti A, et al. , Calibrating the dynamic energy simulation model for an existing building: lessons learned from a collective exercise, Energies. (Basel) 16 (7) (2023) 2979. [Google Scholar]
- [121].Zhang Y, Wang B, Study on indoor thermal environment and energy consumption of traditional dwellings of ethnic minorities in Sichuan plateau, Sci. Rep 15 (1) (2025) 8109. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [122].Connolly CL, et al. , Simulating energy use, indoor temperatures, and utility cost impacts amidst a warming climate in a multi-family housing model, Journal of Urban Health 100 (6) (2023) 1234–1245. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [123].Nyame-Tawiah D, Attuah L, Koranteng C, Determination of the effect of green roofs on indoor temperature by the use of simulation in a tropical landscape, J. Sci. Res. Rep 26 (4) (2020) 12–28. [Google Scholar]
- [124].Liu P, et al. , The use of machine learning to determine moisture recovery in a heat wheel and its impact on indoor moisture, Build. Environ 215 (2022) 108971. [Google Scholar]
- [125].Shrestha A, Shimizu T, Thermal performance assessment of traditional Japanese wooden houses with short-term measurements: machine learning approaches, Journal of Building Engineering 97 (2024) 110954. [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:
Dr Leah Schinasi reports financial support was provided by National Heart Lung and Blood Institute of the National Institutes of Health. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
No data was used for the research described in the article.