Total Environment Research Themes 8 (2023) 100083

Contents lists available at ScienceDirect

Total Environment Research Themes

journal homepage: www.elsevier.com/locate/totert

Development and comparison of adaptive data-driven models for thermal

comfort assessment and control
Giulia Lamberti a, b, *, Roberto Boghetti c, Jérôme H. Kämpf c, Fabio Fantozzi a,
Francesco Leccese a, Giacomo Salvadori a
a
University of Pisa, School of Engineering, Largo Lucio Lazzarino, 56122 Pisa, Italy
b
Institut de Recherche en Constructibilité, Université Paris-Est, ESTP, 28, Avenue du Président Wilson, 94230 Cachan, France
c
Energy Informatics Group, Idiap Research Institute, 1920 Martigny, Switzerland

A R T I C L E I N F O A B S T R A C T

Keywords: Thermal comfort prediction is an important issue, as it can largely influence occupants’ well-being and buildings’
Thermal comfort energy consumption. Nowadays, models used to assess thermal comfort have been increasingly discussed, and a
Data-driven model growing number of data-driven models with several input parameters developed. Although these models allow
Adaptive thermal comfort
reasonably accurate predictions of thermal comfort, using complex algorithms to determine thermal comfort
PMV
might be unsuitable for some use cases, such as quick estimations or real-time control of Heating, Ventilation,
ASHRAE database
and Air Conditioning (HVAC) systems.
In this paper, a data-driven model was developed based on 61710 samples of subjective responses associated
with environmental parameters from field studies available in two ASHRAE databases. Two models resulted from
this analysis, one with higher accuracy and one simplified, which improved the prediction in comparison to other
regression models and PMV.
However, since thermal comfort cannot be conceived as a punctual condition, comfort areas were derived, i.e.,
respective comfort ranges at 90%, 80%, and 70% of thermal acceptability. The result is that the error in the
prediction of the new models is below the 90% acceptable range, which means that the models’ error does not
lead to a reduction in the evaluation of occupant comfort.
Built upon influential parameters, these models enable thermal comfort estimates and occupant-centered
HVAC control. The notion of comfort as a non-fixed state empowers more flexible building management
criteria, reducing energy use while upholding indoor comfort.

TCV Thermal Comfort Vote

List of Symbols Tg Globe temperature (◦ C)
G Griffiths’ constant (◦ C− 1) Top Operative temperature (◦ C)
HVAC Heating Ventilation and Air Conditioning Tout Mean monthly outdoor temperature (◦ C)
Icl Clothing insulation (clo) TSV Thermal Sensation Vote
M Metabolic rate (met) Tr Mean radiant temperature (◦ C)
MAE Mean Absolute Error Va Air velocity (m/s)
PMV Predicted Mean Vote
PD Percentage of Dissatisfied 1. Introduction
PPD Predicted Percentage of Dissatisfied
RH Relative humidity (%) The importance of healthy indoor conditions in buildings has
RMSE Root Mean Square Error become more evident due to increased time spent indoors. Thermal
Ta Air temperature (◦ C) comfort plays a crucial role in people’s well-being, productivity, and
Tcomf Comfort temperature (◦ C) energy consumption (Lamberti, 2020; Zhang et al., 2019). Previous

* Corresponding author.
E-mail address: [email protected] (G. Lamberti).

https://doi.org/10.1016/j.totert.2023.100083
Received 25 July 2023; Received in revised form 31 August 2023; Accepted 5 October 2023
Available online 6 October 2023
2772-8099/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

studies have extensively examined indoor thermal comfort (Enescu, With the purpose of developing a model suitable for real-time HVAC
2017; Fantozzi and Rocca, 2020), with Fanger’s rational and the adap­ system control, two primary challenges were identified. Firstly, the
tive being the primary models used (Djongyang et al., 2010). current adaptive models overlook various factors influencing the heat
Fanger’s rational model (Fanger, 1970) evaluates occupants’ thermal balance between the human body and the environment, thereby
sensations using the Predicted Mean Vote (PMV) index, which is based impacting thermal sensation. Secondly, these adaptive models were
on four environmental (air temperature, relative humidity, mean initially designed to simplify comfort prediction, prioritizing accessi­
radiant temperature, and air velocity) and two individual (metabolic bility for building professionals.
rate and clothing insulation) parameters. The rational model considers Although it has been shown in numerous studies that the PMV often
the heat balance between the human body and the environment and was fails in predicting comfort (Cheung et al., 2019; Humphreys and Nicol,
developed through subjective climate chamber experiments. This model 2002), there is a need to develop models that can accurately predict
is widely used for thermal comfort analysis, particularly in the design occupant comfort using a minimal set of input parameters. These models
phase, but its numerous input parameters make it challenging for real- should be suitable for quick estimations and real-time control.
time control of HVAC systems. This paper aims to fill this gap by developing a data-driven model
To overcome this limitation and account for adaptive processes that:
occurring in buildings, the adaptive model was derived from field
studies (Humphreys et al., 2016). Thermal adaptation, encompassing 1. Is based on a considerable amount of data, to account for adaptive
behavioural, physiological, and psychological aspects (Brager and de processes, which are only detectable from field studies.
Dear, 1998), recognizes occupants as active contributors to comfort 2. Selects and includes the parameters that are most relevant to thermal
creation, not merely passive subjects. Factors like building type (Rupp perception.
et al., 2019), climate (Lamberti et al., 2023), and ventilation mode 3. Allows real-time control and is suitable for use by building practi­
(Kumar, 2022) significantly shape occupants’ thermal perception and tioners, in line with the original adaptive model.
adaptability and should be considered in comfort analysis. The impor­ 4. Presents a satisfying predictive performance, suitable for its use in
tance of adaptation in different settings has become particularly relevant building management.
(Oliveri et al., 2016; Castilla et al., 2018) and passive strategies, based
on occupants’ adaptive capacities, can also improve thermal comfort To overcome the problem that laboratory studies may not represent
conditions (Zinzi et al., 2021). In this case, the only parameters involved accurately occupants’ behaviour, ASHRAE’s databases including data
in the adaptive relationship are the indoor and outdoor temperatures. from real buildings well-spread around the world were used (de Dear,
Although this greatly simplifies the model and makes it more suitable for 1998; Földváry Ličina et al., 2018).
quick estimations and real-time control, it reduces thermal comfort to
the relationship between indoor and outdoor temperatures, without 2. Methodology
considering all the other parameters that are involved in the heat bal­
ance (Fanger and Toftum, 2002). In this section, the methodology used for developing the model is
To consider the issue that the original adaptive models reduce the described (Fig. 1). Rows with parameters Ta, Tr, RH, Va, Icl, M, Tout,
prediction of thermal sensation to the relationship between indoor and PMV, and TSV were extracted from ASHRAE databases. PMV was
outdoor temperature only, machine learning algorithms have been recalculated for comparison with user-perceived TSV. Filtering retained
recently used, leveraging the increasing availability of field data. comfort data (excluding thermal stress), resulting in 62,317 to 61,710
Indeed, data-driven models present high flexibility in input parameters records. Then, the filtered dataset was split into 90% training and 10%
(Xie et al., 2020) so that they can consider aspects such as skin tem­ test sets. Once the baseline was established, a polynomial regression to
perature (Dai et al., 2017), personal characteristics (Li et al., 2017; Lee create two thermal sensation prediction models was performed and the
et al., 2017; Zhao et al., 2014), or occupants’ interaction with control test set was used for validation. Further methodology details are pro­
systems (Kim et al., 2018). Several studies comprised the input param­ vided in the following paragraphs.
eters of PMV including in some cases outdoor temperature or HVAC
operation mode (Jiang and Yao, 2016; Lu et al., 2019). 2.1. Data source and preparation
In particular, they have been applied to the prediction of different
indices, such as 3-points Thermal Sensation Vote (TSV) (Chaudhuri 2.1.1. Description of the databases
et al., 2018; Chaudhuri et al., 2019; Wang et al., 2019), 7- points There are several quality-controlled thermal comfort databases
Thermal Sensation Vote (Jiang and Yao, 2016; Lu et al., 2019; Rana available for scientific studies, such as the ASHRAE RP-884 database
et al., 2013; Wang et al., 2014; Wu et al., 2018; Du et al., 2019), Thermal developed in the 1990 s, with 25,616 samples (de Dear, 1998), the Smart
Preference Vote (Dai et al., 2017; Li et al., 2017; Lee et al., 2017; Kim Control and Thermal Comfort database developed in the 2000 s
et al., 2018), or Thermal Comfort Vote (TCV) (Farhan et al., 2015; (McCartney and Nicol, 2002), with 27,284 samples, and the more recent
Cosma and Simha, 2019). ASHRAE Database II, which has the largest size with 81,968 samples
To build data-driven models, different algorithms were used, which (Földváry Ličina et al., 2018).
showed a median predictive accuracy of 84% with a standard deviation Previously, machine learning algorithms predicted thermal comfort
of around 15% (Xie et al., 2020). While it is true that this predictive using individual databases (Lu et al., 2019; Farhan et al., 2015; Luo
performance is better than that of the PMV (Xie et al., 2020; Cheung et al., 2020; Gao et al., (2019); Zhou et al., 2020). Addressing concerns
et al., 2019) and thus shows how promising the use of these algorithms is from researchers encouraging broader samples (McCartney and Nicol,
for better understanding the parameters that most influence thermal 2002), ASHRAE’s Databases Database II and RP-884 were combined,
comfort, they often require a greater number of input parameters for integrating global field studies with a consistent methodology. Utilizing
predicting thermal comfort indices. both databases offers a comprehensive, extended data range.
In addition, while more complex algorithms have improved thermal Selected model parameters include air temperature (Ta), mean
sensation prediction compared to the PMV model (Xie et al., 2020), their radiant temperature (Tr), relative humidity (RH), air velocity (Va),
effectiveness is limited due to the subjective nature of occupant- clothing insulation (Icl), metabolic rate (M), and mean monthly outdoor
perceived thermal sensations. Variability in individual experiences of temperature (Tout). When the mean radiant temperature was absent but
thermal comfort exists, making precise individual predictions imprac­ the globe temperature (Tg) was available, Tr was derived using the ISO
tical and unnecessary. A pragmatic alternative is to focus on “comfort 7726 standard (ISO 7726, 2001). The six parameters (Ta; Tr, RH, Va, Icl,
zones” rather than specific conditions (Humphreys et al., 2016). M) correspond to those in Fanger’s PMV model, while outdoor

2
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

Fig. 1. Flow chart showing the procedures used for developing the thermal comfort models.

temperature (Tout) reflects adaptability, foundational to the adaptive the predictive model. These error indices can be calculated as follows.
model. ∑n
|pi − ai |
Then, Fanger’s Predicted Mean Vote (PMV) was calculated. The MAE = i=1 (1)
n
database already includes the values of PMV for each entry however, as
it is possible for a database of this size to contain errors (Cheung et al., √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
2019; Humphreys and Nicol, 2002), PMV was recalculated using the √n
√ (pi − ai )2
√
pythermalcomfort Python library (Tartarini and Schiavon, 2020) to RMSE = i=1 (2)
ensure consistency. Samples with missing values of Ta; Tr or Tg, RH, Va, n
Icl, M, Tout, and TSV were excluded from the study. As a result, the ∑n
i=1 (pi − ai )
considered database had a size of 62,317 observations, obtained from Bias = (3)
studies carried out from 1982 to 2016. n
The Thermal Sensation Vote (TSV), which represents occupants’ where pi is the predicted value; ai is the target value, and n is the number
thermal sensation on a 7-points scale (from − 3 cold to + 3 hot), was also of samples.
extracted from the databases for the comparison between the predicted For comparison, the predictive performance of PMV was also
and real thermal sensation. The data originates from peer-reviewed field assessed using these three indices (MAE, RMSE, and Bias).
studies, capturing effective thermal experiences of occupants in real Then, the performance of a linear least squares regression model,
environments. Thermal Sensation Votes (TSV) were gathered using eventually regularized with different values of L1 (Lasso) and L2 (Ridge)
combined questionnaires and instrumental measurements within the penalties, as well as combinations of both (Elastic Net), was calculated
same spatial and temporal context. This approach enabled the correla­ and used as a baseline that the proposed model is expected to improve,
tion of individual thermal responses with concurrent environmental and a primary comparison between this approach and Fanger’s PMV
parameters. Rigorous quality checks and validation procedures were prediction was provided.
applied to all submissions to mitigate potential transmission errors Linear regression is one of the simplest data-driven modelling algo­
(Földváry Ličina et al., 2018). rithms, and it has the advantage that its results can be easily understood
and interpreted. Ridge regression (Hoerl and Kennard, 1970) extends it
2.1.2. Data filtration by introducing the L2 penalty to reduce the variability of the model and
After an initial analysis (see Supplementary Material), data was to mitigate the problem of multicollinearity, as it penalises higher pre­
filtered and values beyond practical ranges were excluded. Acceptability dictors’ coefficients, effectively shrinking them to values that can be
ranges were defined based on thermal comfort standards for each input close, but not equal, to zero. The Lasso regression (Tibshirani, 1996)
feature (ISO 7726, 2001; ISO 7730, 2006). uses a similar approach with the introduction of the L1 penalty term,
For instance, Ta and Tr below 10 ◦ C were omitted due to cold stress which allows the parameters to be effectively reduced to zero. In this
risk. No upper limit was set, considering diverse climates. A wide hu­ way, redundant or not useful predictors are removed, which should lead
midity range of 10–90% was accepted since studies were performed in to a simpler and potentially better model. Finally, the Elastic Net
dry and humid climates. Air velocity ranged from 0.0 to 3.0 m/s, typical regression (Zou and Hastie, 2005) is a method that combines both the L1
for indoor settings. Parameters like metabolic rates below 0.8 Met and L2 penalties used in the Lasso and Ridge regression models.
(sleeping condition) and clothing insulation below 0.0 clo were dis­ The models were built and optimized using Python’s Scikit-learn li­
regarded. PMV and TSV were limited to − 3 to + 3 as per ASHRAE. brary (Pedregosa et al., 2011). A 5x5-fold cross-validated random search
Duplicate samples were removed, resulting in 607 exclusions. was made to find optimal values of the regularization parameters. The
Then, the dataset was split: 90% for training (baseline identification Scikit-learn implementations of these models use the hyperparameters
and model building), and 10% for testing model validation. alpha to control the strength of the total penalty term, while in Elastic
Net the l1_ratio controls the ratio of penalty assigned to the L1 term. The
2.2. Baseline identification range of values tested was 0.05 – 50 for alphas on a logarithmic scale and
0.05–0.95 on the l1_ratio with steps of 0.05. For the Lasso, Ridge, and
The model aims to predict occupants’ thermal sensation (7-point Elastic Net regressions, 100 different regularization penalty parameters
scale) using the identified parameters (Ta, Tr, RH, Va, Icl, M, Tout). After (alphas) were tested and for the Elastic Net, 19 different mixing pa­
identifying influential parameters through field studies, the model aims rameters between Lasso and Ridge (l1_ratio) were examined.
for user-friendliness such as adaptive models. It seeks improved pre­
dictability compared to PMV and linear models, approaching complex 2.3. Polynomial regression
algorithms’ performance.
Model performance is assessed using Mean Absolute Error (MAE), To enhance model performance, an instance selection algorithm was
Root Mean Square Error (RMSE), and Bias. Lower values indicate higher applied to training data, commonly used in classification but less
precision. Bias detects systematic overestimation/underestimation by explored in regression (Song et al., 2017; Arnaiz-González et al., 2016;

3
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

Arnaiz-González et al., 2016). This technique undersamples training Observations were primarily gathered in summer (27778 records),
data, retaining model predictiveness while reducing questionnaire data followed by winter (22922 records), spring (5054 records), and autumn
noise. The instance selection methodology applied was based on clus­ (2542 records). Data were also available for wet (2300 records) and dry
tering, similar to common filter methods used for classification (Olvera- (1114 records) seasons.
López et al., 2010). Noteworthy building types encompass offices (38567 records),
Specifically, complete-linkage hierarchical agglomerative clustering schools (11552 records), residential buildings (7845 records), senior
(Nielsen, 2016) was employed, allowing control over same-cluster centres (449 records), light industrial factories (118 records), and other
element distances. This possibility, coupled with a different pre­ categories (3179 records).
liminary scaling of each feature, results in the ability to impose a Table 2 provides parameter descriptions, including assumed ranges,
maximum difference between extreme values of each feature inside a mean values, and predefined acceptability ranges.
cluster. These constraints were chosen as the limit values beyond which Table 3 presents the parameter correlation matrix. TSV’s weak cor­
the thermal environment is perceived differently by occupants accord­ relations with other inputs reflect personal thermal sensation variability.
ing to the international standard ISO 7726 (ISO 7726, 2001). The limit Notably, TSV displays a notable link with air/mean radiant tempera­
values of thermal environment perception, coinciding with the tures (r = 0.37), underscoring their impact on thermal perception.
maximum distances of each feature allowed in a single cluster, are re­ Additionally, TSV exhibits a significant correlation with clothing insu­
ported in Table 1. lation (r = -0.18), indicating occupants’ adaptive responses. This
Employing this approach, 44,264 clusters emerged from the filtered adaptability is reinforced by Icl’s correlations with Ta (r = -0.44), Tr (r =
database. Instance selection during model training retained central -0.45), and Tout (r = -0.37).
samples from each cluster, refining the dataset. However, TSV’s correlation with metabolic rate is minimal (r = 0.05)
From the filtered database, the parameters were combined in poly­ due to limited sedentary activity variation in the sample. TSV’s
nomials of degree 5th or lower, resulting in a total of 792 polynomials connection with RH is also weak (r = -0.03), despite RH’s broad range
(also called predictor variables or input features in the context of data- (10% to 90%), reflecting the complex role of RH in thermal comfort
driven modelling). To select the polynomials that influence the most (Djamila, 2017). Similarly, TSV’s correlation with air velocity is modest
thermal perception of the building’s occupants, a Lasso regression was (r = 0.05), influenced by low indoor variability. The effect of outdoor
applied to the training set. In fact, Lasso can shrink the independent temperature on TSV is low (r = 0.03), potentially due to database-
variable to 0, effectively removing predictors that are redundant or non- specific building variations.
correlated with the target variable. Before applying the Lasso regression, Noteworthy is the correlation between Icl and Tout, underscoring
each feature was also scaled to a range (0,1) using the MinMaxScaler. outdoor temperature’s significance in thermal sensation, particularly in
While scaling features do not affect the performance of linear least naturally ventilated settings. The correlation matrix underscores occu­
square regression models, this step was necessary to measure the relative pants’ real-world adaptability, evident through clothing insulation’s
importance of the polynomials as a function of the Lasso coefficients. impact on thermal perception.
The Lasso regression was then applied using 5x5-fold cross-validation to
find the optimal value of alpha. 3.2. Model’s baseline
Thermal sensation evaluation was approached as regression, align­
ing with global comfort theories, as thermal comfort analysis typically The model is based on field studies, enabling the evaluation of user
forecasts group thermal perceptions, not individual responses. This is comfort within real environments, which allow to consider thermal
achieved through data binning, averaging sensations under similar adaptation.
conditions. TSV and PMV are expressed as continuous real values within PMV was used as a baseline, following previous studies on data-
the − 3 to + 3 range, facilitating direct comparison. driven models (Xie et al., 2020). Simple regression models were also
used as a reference for predictive ability. It was not possible to consider
3. Results and analysis as baselines other data-driven models in the literature due to their
classification-based approach for TSV evaluation, as discussed earlier.
3.1. Data analysis PMV’s predictive performance was evaluated using three error
indices. It showed a MAE of 0.991, indicating an average error of
ASHRAE’s databases include global studies across diverse climatic approximately one point on ASHRAE’s 7-point scale. The RMSE was
zones, illustrated in Fig. 2 (see Supplementary Material for data distri­ 1.276, slightly higher than the standard deviation of TSV (SDTSV =
bution). The dataset contains 29,236 records from naturally ventilated, 1.268), while the Bias was low at − 0.003, indicating no systematic
23,149 from air-conditioned, and 9325 from mixed-mode buildings. underestimation or overestimation of thermal comfort.
Predominantly, studies were conducted in temperate (Zone C, 38,061 Simple regression models, including linear regression (MAE = 0.873,
records), tropical (Zone A, 11,015 records), dry (Zone B, 7118 records), RMSE = 1.135, and Bias = 0.000), Ridge regression (MAE = 0.873,
and continental (Zone D, 5516 records) climates. RMSE = 1.135, and Bias = 0.000), Lasso (MAE = 0.872, RMSE = 1.145,
and Bias = 0.000), and Elastic Net (MAE = 0.872, RMSE = 1.145, and
Bias = 0.000), showed improved predictive performance compared to
Table 1 PMV across all three evaluation indices.
Limit values of the input parameters used for clustering data. The predictive model developed in the next section should aim to
Parameter Limit Unit further enhance the performance achieved by these simple linear
value models.
Environmental Air temperature (Ta) 0.5 (◦ C)
parameters Relative humidity (RH) 5 (%) 3.3. Prediction of the Thermal Sensation Vote
Mean radiant temperature (Tr) 2 (◦ C)
Air velocity (Va) 0.05 (m/
The relevant environmental and individual parameters were com­
s)
Mean monthly outdoor 1 (◦ C) bined into polynomials, resulting in 792 features for regression models.
temperature (Tout) Lasso regression was then used to select important parameters and
Individual parameters Clothing insulation (Icl) 0.1 (clo) reduce predictors to 37 polynomials, as shown in Table 4.
Metabolic rate (M) 0.1 (Met) The MAE, RMSE, and Bias of the cross-validations, calculated as the
TSV 1 (-)
average value of the validation sets of each fold, were 0.867, 1.121, and

4
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

Fig. 2. Geographical distribution of the data contained in ASHRAE’s databases.

from Lasso regression more practical, it was essential to further simplify

Table 2
the model’s complexity. The objective was to identify the minimum
Description of the Databases’ parameters.
number of polynomials necessary to attain comparable accuracy. This
Parameter Acceptability Range Mean involved repeating curve fitting with progressively larger predictor
range assumed value
subsets until the error approximated that of the full polynomial model.
Air temperature Ta (◦ C) ≥ 10 10.0 – 42.7 24.0 Employing 5x5-fold cross-validation ensured accurate MAE, RMSE, and
Relative humidity RH (%) 10 – 90 10 – 89 47
Bias estimation for each model. The process was halted after the third
Mean radiant temperature Tr (◦ C) ≥ 10 10.0 – 49.5 24.2
Air velocity Va (m/s) 0–3 0.00 – 2.81 0.13
iteration.
Clothing insulation Icl (clo) ≥0 0.03 – 2.87 0.73 Across the combinations, MAE ranged from 0.866 to 0.947, RMSE
Metabolic rate M (Met) ≥ 0.8 0.8 – 4.5 1.2 spanned from 1.128 to 1.261, and Bias fluctuated between − 0.012 and
Mean monthly outdoor – − 24.9 – 35.0 17.5 0.017. Given consistently low Bias, always under the recommended ±
temperature Tout (◦ C)
0.25 threshold (Humphreys and Nicol, 2002), and RMSE below TSV’s
PMV − 3 to + 3 − 3 to + 3 0.13
TSV − 3 to + 3 − 3 to + 3 0.16 standard deviation (SD = 1.307 for the clustered dataset), MAE was
chosen as the criterion for selecting the optimal combination, priori­
tizing model accuracy.
0.001 respectively. The most relevant polynomial is composed of the The best model found, resulting in a MAE of 0.866, a RMSE of 1.128,
product of Ta and Tr, revealing the significance that these two parame­ and a Bias of 0.008 on the cross-validation, was:
ters have on thermal sensation.
For instance, fitting a linear regression with this polynomial only led TSVI = 0.0039Ta ⋅Tr − 3.3259⋅10− 8 Tr2 ⋅RH⋅Tout
2
+ 4.5622⋅10− 7 ⋅Ta3 ⋅RH⋅M
(4)
to a MAE, RMSE, and Bias of 0.886, 1.165, and 0.003 respectively. Fig. 3 − 2.1152
reports the Lasso coefficients of the different features in the subset,
which show their relative importance in determining the thermal This equation suggests that the combination of the most important
sensation. In blue, the features used for the model of Eq. (4), while in features that determine the thermal sensation includes Ta, Tr, RH, Tout,
light blue the features used for the model of Eq. (5) are reported. and M.
The feature comprising Ta and Tr stands out as the most important, However, this model includes parameters that are difficult to control
confirming their influence on the perception of the thermal environment in real-time, such as mean radiant temperature or metabolic rate, which
in the adaptive model. However, the effects of relative humidity, out­ is typically estimated using tables based on the activity performed for
door temperature, and metabolic rate included in Eq. (4) should not be global thermal comfort assessment (Cheung et al., 2019). To simplify
neglected. further the equation another combination was considered, which had
To make the utilization of the 37 terms in the polynomial equations slightly higher MAE and RMSE, but lower Bias (MAE = 0.872, RMSE =

Table 3
Correlation matrix expressing the correlation coefficient (r) between the different parameters.
TSV Ta RH Va Tr M Icl Tout

TSV 1 0.37 − 0.03 0.05 0.37 0.05 − 0.18 0.03

Ta 0.37 1 0.16 0.39 0.93 − 0.07 − 0.44 0.53
RH − 0.03 0.16 1 0.22 0.19 − 0.08 − 0.30 0.50
Va 0.05 0.39 0.22 1 0.36 − 0.06 − 0.16 0.36
Tr 0.37 0.93 0.19 0.36 1 − 0.05 − 0.45 0.52
M 0.05 − 0.07 − 0.08 − 0.06 − 0.05 1 − 0.04 − 0.10
Icl − 0.18 − 0.44 − 0.30 − 0.16 − 0.45 − 0.04 1 − 0.37
Tout 0.03 0.53 0.50 0.36 0.52 − 0.10 − 0.37 1

5
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

Table 4 variations in environmental parameters were investigated, focusing on

Features selected after the application of the Lasso regression. Bias as the error indicator (Fig. 4). For comparison purposes, PMV was
ID Features ID Features included, aligning with prior research (Xie et al., 2020).
The models’ validity was evaluated by establishing accurate pre­
1 Ta • Tr 20 Tr • RH • Tout
2 Tr • RH3 • Icl 21 Ta • Icl • T2out diction ranges for both environmental and individual parameters. Fig. 4
3 T5out 22 Ta • RH2 • I2cl illustrates the difference between predicted and actual thermal sensa­
4 T2r 23 RH • V2a • M • Tout tion values across diverse parameters. This assessment was carried out
5 T3a • Tr • RH 24 RH2 • M on the test set, confirming the models’ reliability and setting their
6 RH • I4cl 25 RH3 • M
7 Tout 26 Va • T4out
operational boundaries. The models were calculated based on an
8 Tr • T4out 27 T2a • Tr • RH • M acceptable Bias range of ± 0.25 points on the TSV scale (Humphreys and
9 Icl • T4out 28 T2a • RH • Va • Icl Nicol, 2002). In Fig. 4, it’s evident that despite significant response
10 T2a • RH • M • Tout 29 T2r • RH • T2out variability, the new models exhibit a tendency to decrease the Bias be­
11 Ta • M • Icl • T2out 30 M • Tout
tween predicted and observed thermal sensations compared to PMV.
12 RH3 31 Ta • Va • Icl
13 V2a • T3out 32 Va The new models perform well in predicting sensations within indoor
14 I2cl 33 T2a • RH2 • Tout air temperatures of 10 ◦ C to 35 ◦ C, albeit overestimating for temper­
15 Ta • Icl • Tout 34 RH • Va • M3 atures beyond 35 ◦ C. Unlike the original PMV model, valid from 10 ◦ C to
16 Ta • M 35 Va • I2cl • T2out 30 ◦ C, the new models improve and extend predictive capacity.
17 M • Icl 36 Ta • Va • M2 • Icl
18 RH2 • Tout 37 T3a • RH • M
Regarding models in Eq. (4) and (5), distinctions mainly emerge in
19 Ta 10–15 ◦ C and 35–40 ◦ C ranges, often beyond moderate comfort,
requiring specialized indices for heat or cold stress. For accurate thermal
sensation forecasts using the new models, maintaining air temperature
1.136, Bias = 0.001). With this equation, it is possible to predict thermal within 15–35 ◦ C is advised. The range could potentially expand to
sensation with few environmental parameters that are easy to measure, 10–35 ◦ C with minor precision reduction (Table 6). Beyond these limits,
maintaining the predictive performance close to the previous one. model performance might decrease, elevating heat or cold stress risks.
TSVII = − 0.0228Tout − 1.9492⋅10− 6 RH 2 ⋅Tout + 0.1867Ta − 3.8609 (5) Bias remained consistent across various relative humidity ranges
(Fig. 4), indicating occupants’ modest sensitivity to humidity modifi­
Table 5 summarizes the indices of error for MAE, RMSE, and Bias for cations. Overall, the proposed models slightly surpass PMV in predictive
simple regression models (linear, Lasso, Ridge, and Elastic Net), PMV, capability, with median Bias often within acceptable thresholds.
and the two developed models. It can be noticed that the best predictive Particularly, models Eq. (4) and (5) display greater enhancements over
performance is achieved with the Model of Eq. (4). PMV for relative humidity values surpassing 60%. While not immedi­
It can be noticed that the correlation matrix reveals a consistent in­ ately perceptible, relative humidity considerably impacts long-term
verse relationship between TSV and Icl. This means that an adaptive
tendency to adjust clothing’s thermal insulation based on perceived
warmth or coldness, and vice versa to experience warmth or coldness Table 5
while donning more or less insulated clothing. However, in the derived MAE, RMSE and Bias for simple regression models, PMV and the two developed
models.
models, Icl is not included in the equations. This outcome, rigorously
derived, aligns with adaptive models, which recognise the adaptive Model MAE RMSE Bias
equation in the relationship between outdoor and indoor temperatures. Linear regression 0.873 1.135 0.000
Lasso regression 0.872 1.135 0.000
3.4. Validation and boundaries of the model Ridge regression 0.873 1.135 0.000
Elastic Net regression 0.872 1.135 0.000
PMV 0.991 1.276 − 0.003
The databases employed in this analysis encompass typical envi­ Model Eq (4) 0.866 1.128 0.008
ronments, which implies that the model’s predictions might deviate Model Eq (5) 0.872 1.136 0.001
under extreme conditions. To assess the models’ performance, the

Fig. 3. Lasso coefficients associated to the different features.

6
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

Fig. 4. Bias between the predicted and real thermal sensation against different parameters. On the x-axis the range of the parameters is reported, and the number of
samples is reported below them. In grey Bias in PMV, in blue bias in model of Eq. (4), and in light blue Bias in model of Eq. (5). (For interpretation of the references to
colour in this figure legend, the reader is referred to the web version of this article.)

7
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

Table 6 overestimation or underestimation, Bias was calculated for each pre­

Boundaries of PMV, models of Eq. (4) and (5). dicted range. PMV and models Eq. (4) and (5) cover the − 3 to + 3 TSV
Parameter PMV Model Eq. (4) Model Eq. (5) range, shown on the x-axis of Fig. 4. This allows measuring Bias between
predicted and observed sensations when the model predicts specific
Air temperature Ta ( C)
◦
10.0 – (10*)15.0 – (10*)15.0 –
30.0 35.0 35.0 sensations, illustrating model performance on the 7-point ASHRAE
Relative humidity RH (%) – 10.0 – 80.0 10.0 – 80.0 scale. Overall, the two new models demonstrate lower errors than PMV
Partial vapour pressure (Pa) 0 – 2700 – – in practical conditions, despite being simplified and requiring fewer
Mean radiant temperature Tr (◦ C) 10.0 – 15.0 – 35.0 15.0 – 35.0 input parameters. On average, these models tend to reduce PMV’s
40.0
Air velocity Va (m/s) 0–1 0.0 – 0.5 0.0 – 0.5
overestimation/underestimation for TSV < -1 or TSV > 1. However, it’s
(1.5*) (1.5*) important to acknowledge that values close to ± 3 might indicate heat
Clothing insulation Icl (clo) 0.0 – 2.0 0.0 – 1.6 0.0 – 1.6 or cold stress situations, demanding careful utilization of thermal com­
(2.4*) (2.4*) fort indices.
Metabolic rate M (Met) 0.8 – 4.0 0.8 – 2.0 0.8 – 2.0
Mean monthly outdoor temperature – − 15.0 – 35.0 − 5.0 – 35.0
Tout (◦ C) 4. Discussion
Model prediction − 3 to + 3 − 3 to + 3 − 3 to + 3
4.1. Definition of the “comfort areas”
*Extended value reduces the precision.

Enhancing predictive accuracy is crucial, yet not the sole determi­

health (Taylor, 2020). The models suggest suitability within 10–80%
nant of efficacy. Comfort is better understood as a range rather than a
relative humidity (Table 6), but maintaining values between 40% and
fixed state, in line with adaptive principles (Lamberti et al., 2023). This
60% is advisable.
section compares the “comfort zones” established by conventional
Bias trends for mean radiant temperature parallel those of air
adaptive models with those originating from the present models.
temperature (Fig. 4). Model Eq. (4) is less accurate within 10–15 ◦ C,
To derive the thermal comfort range it was assumed a relative hu­
excelling in higher mean radiant temperatures (>30 ◦ C), where PMV
midity of 50%, a metabolic rate of 1.2 met, and a uniform environment
tends to overestimate sensations. However, the new models tend to
(Ta = Tr = Top, where Top is the operative temperature), as recom­
overestimate thermal sensations above 35 ◦ C. For optimal model per­
mended by international standards (EN 16798-1, 2019). These values
formance, use within 15–35 ◦ C or 10–35 ◦ C with slight precision
were chosen as they are typical of many indoor environments, but the
reduction (Table 6) is recommended, still surpassing PMV. Beyond these
model can accommodate different conditions within the defined
ranges, model predictive capability wanes, heightening heat or cold
boundaries in Table 6.
stress likelihood.
Using these assumptions, the comfort temperatures were obtained
Indoor environments typically feature limited air velocity, with
for the models of Eq. (4) and (5), and the relationship between outdoor
most samples within 0.0 to 0.5 m/s range (Fig. 4). This accounts for Bias
temperature (Tout) and comfort temperature (Tcomf) was analysed
distribution’s variability in this range, although models exhibit reduced
through regression analysis for Tout ranging from 10 to 35 ◦ C. This
variability. These models often overestimate sensations in 0.5–1.0 m/s
relationship is depicted by the solid line in Fig. 5. There is to notice that,
and 1.5–2.0 m/s ranges. The 2.0–2.5 m/s range has few samples due to
based on adaptive thermal comfort research, achieving comfort for
the rarity of such high velocities in daily life. Accuracy within 0.5–1.0
operative temperatures up to 35 ◦ C is possible only in certain types of
m/s slightly falls below ± 0.25 range. Generally, predictions within
climates (Rawal et al., 2022), which were included in the development
0.0–0.5 m/s can be deemed accurate, extending to 0.0–1.5 m/s with
of the current models.
minimal precision reduction (Table 6).
For models of Eq. (4) and (5) the acceptability ranges were calculated
Clothing insulation significantly influences occupants’ thermal
for 90% acceptability (Category I, dotted line), with the assumption that
sensation. The new models notably enhance predictive performance,
TSV=±0.5 and a Percentage of Dissatisfied (PD) equal to 10%, for 80%
particularly within the 1.2–2.0 clo range. Higher insulation values (>2
acceptability (Category II, dashed line) assuming TSV=±0.8 and PD =
clo) yield relatively fewer samples (Fig. 4) as occupants indoors typically
20%, and for 70% acceptability (Category III, dashed and dotted line)
avoid excessively warm clothing. As such, applying the new models
assuming TSV=±1.1 and PD = 30% (ISO 7730, 2006).
within the 0–1.6 clo range is recommended, extendable up to 2.4 clo
Table 7 shows the ranges that define the comfort areas for Eqs. (4)
with slightly increased uncertainty (Table 6).
and (5). It will be necessary to add and subtract respectively ΔUP and
Metabolic rate strongly impacts thermal sensation, but accurate
ΔLOW to the above equations to obtain the desired comfort ranges. It can
estimation can be challenging, leading to consistent bias. Bias remains
be observed that the comfort temperature ranges for Models of Eq. (4)
relatively low within 0.8–1.5 Met range (Fig. 4). Model Eq. (4) improves
and (5) are larger than those given by the EN 16798–1 standard (EN,
predictions even for higher Met values (1.5–2.5 Met), while model Eq.
2019).
(5) underestimates sensations in 1.5–2.0 Met range. All models over­
This means that occupants have a significant adaptive capacity and
estimate sensations for high metabolic rates (3.5–4.0 Met). Notably, the
can be in comfort situations over a wide temperature range. It should be
database includes limited samples with Met values > 2.0 Met, as typical
noted, however, that it is appropriate to refer mainly to Categories I and
indoor environments (offices, residential buildings, schools) rarely
II, as Category III leads to environmental conditions that may be toler­
exhibit such high activity levels. Hence, applying the two models within
able within certain limits, but plausibly not conditions of comfort for a
0.8–2 Met range is recommended (Table 6).
large proportion of individuals.
Mean monthly outdoor temperature is not a PMV input, but its
The relationship between Tout and Tcomf for the model of Eq. (4) is
inclusion in models of Eq. (4) and (5) accounts for its significance in
represented by a curve (Fig. 5). It is possible to notice that the comfort
thermal adaptation. Very low temperatures (-25 ◦ C to − 15 ◦ C) result in
temperature increases significantly with the outdoor temperature, while
poor predictive performance for all models, while accuracy improves for
it remains stable for lower Tout. This is because at lower temperatures
higher temperatures (especially − 5◦ C to 35 ◦ C) (Fig. 4). Limited sample
heating systems are usually provided while cooling systems not always
size exists within the − 15 ◦ C to − 5◦ C range, with model Eq. (5) exhib­
are.
iting a tendency to overestimate sensations. Consequently, model Eq. (5)
Thus, people present increased adaptation to higher outdoor tem­
should be restricted to − 5◦ C to 35 ◦ C range, while model Eq. (4) can
peratures, which can have significant implications also on energy con­
potentially extend to lower temperatures down to − 15 ◦ C (Table 6).
sumption. This trend is in line with adaptive thermal comfort, which
For precise thermal sensation prediction, free from systematic
shows that below a certain outdoor temperature, the comfort

8
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

Fig. 5. Relationship between mean Tout and Tcomf for the model of Eq. (4) (a), and model of Eq. (5) (b), assuming RH = 50%, M = 1.2 met, and Ta = Tr = Top. Legend:
comfort temperature (solid line), 90% acceptability range (dotted line), 80% acceptability range (dashed line), and 70% acceptability range (dashed and dotted line).

Table 7
Upper (ΔUP), lower (ΔUP), and total (ΔTOT) ranges of comfort temperatures for the models of Eq. (4), (5), and EN16798-1 (EN, 2019) standard.
Model Eq. (4) Model Eq. (5) EN 16798–1
ΔUP ΔLOW ΔTOT ΔUP ΔLOW ΔTOT ΔUP ΔLOW ΔTOT
(◦ C) (◦ C) (◦ C) (◦ C) (◦ C) (◦ C) (◦ C) (◦ C) (◦ C)

Category I 2.3 2.6 4.9 2.6 2.7 5.3 2.0 3.0 5.0
Category II 3.5 4.4 7.9 4.2 4.3 8.5 3.0 4.0 7.0
Category III 4.7 6.5 11.2 5.9 5.9 11.8 4.0 5.0 9.0

temperature remains constant (Humphreys et al., 2016). The comfort (◦ C− 1), assumed equal to 0.50 (Humphreys et al., 2013). Although
temperatures reached with the model of Eq. (4) are similar to the ones recent studies show that the Griffiths constant can also be a variable, this
given by EN 16798–1 standard (EN, 2019) but with an increased comfort value was chosen because it was estimated from the data available in the
area and a non-linear relationship between Tout and Tcomf. ASHRAE databases, which were also used for the development of the
The model of Eq. (5), like the original adaptive model (EN, 2019) new models.
presents instead a linear relationship between Tout and Tcomf (Fig. 5) but Assuming that Tcomf is equal to Top (i.e. occupants in comfortable
the adaptation is lower and the slope is less steep. Concerning the conditions), a MAE of 0.866 means that the model predicts a TSV equal
comfort temperatures, they result lower than the ones given by stan­ to this value and thus an error in Tcomf of 1.73 ◦ C. In the case of the
dards (EN, 2019) especially for high Tout, while the range of comfort in model in Eq. (5) (MAE = 0.872), the error in the prediction of Tcomf is
model of Eq. (5) is again wider than the original adaptive model. 1.74 ◦ C.
In comparison to the adaptive model of EN 16798–1 (EN, 2019), the Comparing these results with Table 7, the error of about 1.7 ◦ C is
two models are favourable as they can be used not only for the fixed below the 90% acceptable range corresponding to Category I. This
environmental and individual parameters (low metabolic rate, fixed means that the error of the new models is, on a practical level, accept­
range of humidity, etc.) but within the acceptability range discussed in able, as it does not actually lead to a reduction in users’ comfort.
the previous section. Furthermore, the increased range of comfort In general, these effective models are valuable for thermal comfort
temperatures has implications on energy consumption, even if the use of prediction. When real-time data on mean radiant temperature and
the model of Eq. (4) is more appropriate to this aim. metabolic rate are available, Eq. (4) is recommended. Alternatively, Eq.
(5) predicts accurate thermal sensation using only indoor air tempera­
ture, relative humidity, and monthly mean outdoor temperature.
4.2. Practical applications of the new models
4.3. Limitations and future studies
Since these models were developed for actual use in building man­
agement, it is necessary to understand how much the possible error in Although the two new models were developed using a comprehen­
the prediction of the thermal sensation of the two models (Table 5) can sive database, they present some limitations. Indeed, data were mostly
affect thermal comfort. To this aim, the temperature variation associated collected in the comfort area of Thermal Sensation Vote (TSV) and few
with an MAE of 0.866 (Model of Eq. (4)) and 0.872 (Model of Eq. (5)) cases reported TSV out of the range ± 2. For this reason, models of Eq.
must be investigated. (4) and (5) perform better when discomfort is not too high, and in the
To quantify the error in the prediction of TSV in terms of tempera­ case of extreme conditions indices for heat or cold stress should be
ture, Griffiths’ method was used. (Griffiths, 1990). According to this adopted. Future studies should select a wider range of input parameters,
method, the comfort temperature can be calculated as follows (Eq. (6)): especially for the metabolic rate that remained in a restricted range,
TSV despite a large amount of data.
Tcomf = Top + (6) Furthermore, for the development of the model, a database
G
comprehensive of a wide range of building types, operation modes, and
where Tcomf is the comfort temperature, and G is the Griffiths’ constant climates was used. However, the perception of thermal comfort can be

9
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

also influenced by the background and the level of expectations of the Funding
users, which can be a function of the previously mentioned issues.
Although predictive ability has been shown to be satisfactory, future This research was partially funded by University of Pisa, within the
studies should focus on the analysis of different building types, opera­ program PRA 2022, competitive call for the financial support to multi­
tion modes, expectation levels, etc. disciplinary research projects developed at the university Pisa (ID
Finally, future research should be focused on assessing not only the research project PRA_2022_70).
thermal sensation, but also preference and acceptability, and to consider
possible diversities among them.
Despite these limitations, the proposed models constitute a useful Declaration of Competing Interest
and user-friendly approach that improves the prediction if compared to
previous thermal comfort models and can be used by researchers and The authors declare that they have no known competing financial
practitioners to assess thermal comfort and possibly reduce the associ­ interests or personal relationships that could have appeared to influence
ated energy consumption. the work reported in this paper.

5. Conclusions Data availability

In this study, two original data-driven models to predict the thermal Data will be made available on request.
sensation of buildings’ occupants were developed, using a wide sample
of field studies from ASHRAE databases, which allow to account for Appendix A. Supplementary data
thermal adaptation. The purpose was to develop a data-driven model
that (i) includes the adaptive processes, (ii) selects and includes the most Supplementary data to this article can be found online at https://doi.
relevant parameters in thermal perception, (iii) can be practically used org/10.1016/j.totert.2023.100083.
for real-time control, and (iv) presents a good predictive performance.
The first point was addressed by developing the models on large References
databases based on field studies, which allows to account for adaptation.
The second point was fulfilled by having included in the analysis all Arnaiz-González, Á., Díez-Pastor, J.F., Rodríguez, J.J., García-Osorio, C., 2016. Instance
selection for regression: Adapting DROP. Neurocomputing 201, 66–81. https://doi.
the parameters that are considered to be involved in the perception of org/10.1016/j.neucom.2016.04.003.
the thermal environment, which are air temperature (Ta), mean radiant Arnaiz-González, Á., Díez-Pastor, J.F., Rodríguez, J.J., García-Osorio, C.I., 2016. Instance
temperature (Tr), relative humidity (RH), air velocity (Va), clothing selection for regression by discretization. Expert Systems with Applications 54,
340–350. https://doi.org/10.1016/j.eswa.2015.12.046.
insulation (Icl), metabolic rate (M) and outdoor monthly air temperature Brager, G.S., de Dear, R.J., 1998. thermal adaptation in the built environment: a
(Tout). These parameters were combined in polynomials and used to fit a literature review. Energy and Buildings 27, 83–96. https://doi.org/10.1016/S0378-
Lasso regression to define which subset of combinations has the highest 7788(97)00053-4.
Castilla, N., Llinares, C., Bisegna, F., Blanca-Giménez, V., 2018. Affective evaluation of
predictive power concerning the thermal sensation in real case studies the luminous environment in university classrooms. Journal of Environmental
and therefore select the parameters that most influence it. Psychology 58, 52–62. https://doi.org/10.1016/j.jenvp.2018.07.010.
The third point was considered by providing easy-to-use equations to Chaudhuri, T., Zhai, D., Soh, Y.C., Li, H., Xie, L., 2018. Thermal comfort prediction using
normalized skin temperature in a uniform built environment. Energy and Buildings
define the thermal sensation in buildings, namely Eq. (4), and (5). 159, 426–440. https://doi.org/10.1016/j.enbuild.2017.10.098.
Finally, the new models improved the predictive performance when Chaudhuri, T., Soh, Y.C., Li, H., Xie, L., 2019. A feedforward neural network based
compared to simpler regression models or the typically used PMV, as indoor-climate control framework for thermal comfort and energy saving in
buildings. Applied Energy 248, 44–53. https://doi.org/10.1016/j.
shown in Table 5. In particular, the predictive performance of the model
apenergy.2019.04.065.
of Eq. (5) is slightly reduced (MAE = 0.872, Bias = 0.001, RMSE = Cheung, T., Schiavon, S., Parkinson, T., Li, P., Brager, G., 2019. Analysis of the accuracy
1.136) if compared to Eq. (4) (MAE = 0.866, RMSE = 1.128, Bias = on PMV – PPD model using the ASHRAE global thermal comfort database II. Building
0.008), but it is significantly lower than the PMV. and Environment 153, 205–217. https://doi.org/10.1016/j.buildenv.2019.01.055.
Cosma, A.C., Simha, R., 2019. Machine learning method for real-time non-invasive
The two new models present an increased performance if compared prediction of individual thermal preference in transient conditions. Building and
to simple linear models and to Fanger’s PMV, providing a model that can Environment 148, 372–383. https://doi.org/10.1016/j.buildenv.2018.11.017.
be easily applied by practitioners and that takes into account influential Dai, C., Zhang, H., Arens, E., Lian, Z., 2017. Machine learning approaches to predict
thermal demands using skin temperatures: Steady-state conditions. Building and
parameters in the perception of comfort. Furthermore, in comparison to Environment 114, 1–10. https://doi.org/10.1016/j.buildenv.2016.12.005.
the original adaptive model, these new models can also be used for de Dear, R., 1998. Global database of thermal comfort field experiments. ASHRAE
variable environmental and individual parameters. Transactions 104, 1141–1152.
Djamila, H., 2017. Indoor thermal comfort predictions: Selected issues and trends.
With these models it was also possible to define “comfort areas”, Renewable and Sustainable Energy Reviews 74, 569–580. https://doi.org/10.1016/
taking into account an occupant acceptability of 90%, 80%, and 70% j.rser.2017.02.076.
(Categories I, II, III respectively). In addition, by deriving the accept­ Djongyang, N., Tchinda, R., Njomo, D., 2010. Thermal comfort: A review paper.
Renewable and Sustainable Energy Reviews 14, 2626–2640. https://doi.org/
ability ranges for conditions typically found indoors (RH = 50%, M = 10.1016/j.rser.2010.07.040.
1.2 met, Ta = Tr = Top), it was shown that the comfort ranges found for Du, C., Li, B., Liu, H., Ji, Y., Yao, R., Yu, W., 2019. Quantification of personal thermal
the new models are wider than the adaptive ones, especially for Cate­ comfort with localized airflow system based on sensitivity analysis and classification
tree model. Energy and Buildings 194, 1–11. https://doi.org/10.1016/j.
gories II and III (by about 1 ◦ C for Category II and 2 ◦ C for Category III),
enbuild.2019.04.010.
resulting in energy savings in their use. EN 16798-1, Energy performance of buildings - Ventilation for buildings - Part 1: Indoor
In conclusion, leveraging extensive comfort databases enabled the environmental input parameters for design and assessment of energy performance of
development of two improved models to predict thermal sensation, ac­ buildings addressing indoor air quality, thermal environment, lighting and acoustics,
2019.
counting for multiple influencing factors, that surpass the accuracy of Enescu, D., 2017. A review of thermal comfort models and indicators for indoor
commonly employed ones. These models could be integrated into an environments. Renewable and Sustainable Energy Reviews 79, 1353–1379. https://
updated database version, offering calculated comfort indices aligned doi.org/10.1016/j.rser.2017.05.175.
Fanger, P.O., 1970. Thermal comfort. Analysis and applications in environmental
with respondents’ subjective thermal perception. engineering. Danish Technical Press, McGraw-Hill, Copenhagen.
Fanger, P., Toftum, J., 2002. Extension of the PMV model to non-air-conditioned
buildings in warm climates. Spec. Issue Therm. Comf. Stand. 34, 533–536. https://
doi.org/10.1016/S0378-7788(02)00003-8.

10
G. Lamberti et al. Total Environment Research Themes 8 (2023) 100083

Fantozzi, F., Rocca, M., 2020. An extensive collection of evaluation indicators to assess McCartney, K.J., Nicol, J.F., 2002. Developing an adaptive control algorithm for europe.
occupants’ health and comfort in indoor environment. Atmosphere. 11, 90. https:// Spec. Issue Therm. Comf. Stand. 34, 623–635. https://doi.org/10.1016/S0378-7788
doi.org/10.3390/atmos11010090. (02)00013-0.
Farhan, A.A., Pattipati, K., Wang, B., Luh, P., Predicting individual thermal comfort using Nielsen, F., 2016. Introduction to HPC with MPI for data science. In: Hierarchical Clust.
machine learning algorithms, in,, 2015. IEEE int. Conf. Autom. Sci. Eng. CASE 2015, Springer International Publishing, Cham, pp. 195–211.
708–713. https://doi.org/10.1109/CoASE.2015.7294164. Oliveri, M., Pennisi, S., Scaccianoce, G., Vaccaro, V., Bisegna, F., The relavance of indoor
Földváry Ličina, V., Cheung, T., Zhang, H., de Dear, R., Parkinson, T., Arens, E., Chun, C., comfort in the process of prisoners’ rehabilitation: A case study, in,, 2016. IEEE 16th
Schiavon, S., Luo, M., Brager, G., Li, P., Kaam, S., Adebamowo, M.A., Andamon, M. int. Conf. Environ. Electr. Eng. EEEIC 2016, 1–6. https://doi.org/10.1109/
M., Babich, F., Bouden, C., Bukovianska, H., Candido, C., Cao, B., Carlucci, S., EEEIC.2016.7555673.
Cheong, D.K.W., Choi, J.-H., Cook, M., Cropper, P., Deuble, M., Heidari, S., Olvera-López, J.A., Carrasco-Ochoa, J.A., Martínez-Trinidad, J.F., Kittler, J., 2010.
Indraganti, M., Jin, Q., Kim, H., Kim, J., Konis, K., Singh, M.K., Kwok, A., A review of instance selection methods. Artificial Intelligence Review 34, 133–143.
Lamberts, R., Loveday, D., Langevin, J., Manu, S., Moosmann, C., Nicol, F., Ooka, R., https://doi.org/10.1007/s10462-010-9165-y.
Oseland, N.A., Pagliano, L., Petráš, D., Rawal, R., Romero, R., Rijal, H.B., Sekhar, C., Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Schweiker, M., Tartarini, F., Tanabe, S., Tham, K.W., Teli, D., Toftum, J., Toledo, L., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J.T., Passos, A.,
Tsuzuki, K., De Vecchi, R., Wagner, A., Wang, Z., Wallbaum, H., Webb, L., Yang, L., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, É., 2011. Scikit-learn: Machine
Zhu, Y., Zhai, Y., Zhang, Y., Zhou, X., 2018. Development of the ASHRAE global learning in python. Journal of Machine Learning Research 12, 2825–2830.
thermal comfort database II. Building and Environment 142, 502–512. https://doi. Rana, R., Kusy, B., Jurdak, R., Wall, J., Hu, W., 2013. Feasibility analysis of using
org/10.1016/j.buildenv.2018.06.022. humidex as an indoor thermal comfort predictor. Energy and Buildings 64, 17–25.
G. Gao, J. Li, Y. Wen, Energy-Efficient Thermal Comfort Control in Smart Buildings via https://doi.org/10.1016/j.enbuild.2013.04.019.
Deep Reinforcement Learning, (2019). arXiv:1901.04693. Rawal, R., Shukla, Y., Vardhan, V., Asrani, S., Schweiker, M., de Dear, R., Garg, V.,
Griffiths, I., 1990. Thermal comfort studies in buildings with passive solar features. Field Mathur, J., Prakash, S., Diddi, S., Ranjan, S.V., Siddiqui, A.N., Somani, G., 2022.
studies. Adaptive thermal comfort model based on field studies in five climate zones across
Hoerl, A.E., Kennard, R.W., 1970. Ridge regression: Biased estimation for nonorthogonal india. Building and Environment 219, 109187. https://doi.org/10.1016/j.
problems. Technometrics 12, 55–67. https://doi.org/10.1080/ buildenv.2022.109187.
00401706.1970.10488634. Rupp, R.F., Kim, J., Ghisi, E., de Dear, R., 2019. thermal sensitivity of occupants in
M. Humphreys, F. Nicol, R. S., Adaptive Thermal Comfort: Foundations and Analysis, different building typologies: the griffiths constant is a variable. Energy and
Routledge, London, 2016. https://doi.org/10.4324/9781315765815. Buildings 200, 11–20. https://doi.org/10.1016/j.enbuild.2019.07.048.
Humphreys, M., Nicol, F., 2002. The validity of ISO-PMV for predicting comfort votes in Song, Y., Liang, J., Lu, J., Zhao, X., 2017. An efficient instance selection algorithm for k
every-day thermal environments. Energy and Buildings 34, 667–684. https://doi. nearest neighbor regression. Neurocomputing 251, 26–34. https://doi.org/10.1016/
org/10.1016/S0378-7788(02)00018-X. j.neucom.2017.04.018.
Humphreys, M.A., Rijal, H.B., Nicol, J.F., 2013. Updating the adaptive relation between Tartarini, F., Schiavon, S., 2020. Pythermalcomfort: A python package for thermal
climate and comfort indoors; new insights and an extended database. Building and comfort research. SoftwareX. 12, 100578 https://doi.org/10.1016/j.
Environment 63, 40–55. https://doi.org/10.1016/j.buildenv.2013.01.024. softx.2020.100578.
ISO 7726, Ergonomics of the thermal environment – Instruments for measuring physical S. Taylor, Putting People First: The Healing Power of Indoor Air, (2020). https://www.
quantities, 2001. ashrae.org/file%20library/technical%20resources/covid-19/seminar-21_
ISO 7730, Ergonomics of the thermal environment - Analytical determination and presentation-1.pdf.
interpretation of thermal comfort using calculation of the PMV and PPD indices and Tibshirani, R., 1996. Regression shrinkage and selection via the lasso. J. R. Stat. Soc. Ser.
local thermal comfort criteria, 2006. B-Methodol. 58, 267–288.
Jiang, L., Yao, R., 2016. Modelling personal thermal sensations using c-Support vector Wang, Z., Li, A., Ren, J., He, Y., 2014. thermal adaptation and thermal environment in
classification (C-SVC) algorithm. Building and Environment 99, 98–106. https://doi. university classrooms and offices in harbin. Energy and Buildings 77, 192–196.
org/10.1016/j.buildenv.2016.01.022. https://doi.org/10.1016/j.enbuild.2014.03.054.
Kim, J., Zhou, Y., Schiavon, S., Raftery, P., Brager, G., 2018. Personal comfort models: Wang, Z., Yu, H., Luo, M., Wang, Z., Zhang, H., Jiao, Y., 2019. Predicting older people’s
Predicting individuals’ thermal preference using occupant heating and cooling thermal sensation in building environment through a machine learning approach:
behavior and machine learning. Building and Environment 129, 96–106. https://doi. Modelling, interpretation, and application. Building and Environment 161, 106231.
org/10.1016/j.buildenv.2017.12.011. https://doi.org/10.1016/j.buildenv.2019.106231.
Kumar, S., 2022. Subject’s thermal adaptation in different built environments: an Wu, Z., Li, N., Peng, J., Cui, H., Liu, P., Li, H., Li, X., 2018. Using an ensemble machine
analysis of updated metadata-base of thermal comfort data in india. J. Build. Eng. learning methodology-Bagging to predict occupants’ thermal comfort in buildings.
46, 103844 https://doi.org/10.1016/j.jobe.2021.103844. Energy and Buildings 173, 117–127. https://doi.org/10.1016/j.
Lamberti, G., Leccese, F., Salvadori, G., Contrada, F., Kindinis, A., 2023. investigating the enbuild.2018.05.031.
effects of climate on thermal adaptation: A comparative field study in naturally Xie, J., Li, H., Li, C., Zhang, J., Luo, M., 2020. Review on occupant-centric thermal
ventilated university classrooms. Energy and Buildings 294, 113227. https://doi. comfort sensing, predicting, and controlling. Energy and Buildings 226, 110392.
org/10.1016/j.enbuild.2023.113227. https://doi.org/10.1016/j.enbuild.2020.110392.
Lamberti, G., Thermal comfort in the built environment: current solutions and future Zhang, F., de Dear, R., Hancock, P., 2019. Effects of moderate thermal environments on
expectations, in,, 2020. IEEE int. Conf. EEEIC ICPS Eur. 2020, 1–6. https://doi.org/ cognitive performance: A multidisciplinary review. Applied Energy 236, 760–777.
10.1109/EEEIC/ICPSEurope49358.2020.9160558. https://doi.org/10.1016/j.apenergy.2018.12.005.
Lee, S., Bilionis, I., Karava, P., Tzempelikos, A., 2017. A bayesian approach for Zhao, Q., Zhao, Y., Wang, F., Wang, J., Jiang, Y., Zhang, F., 2014. A data-driven method
probabilistic classification and inference of occupant thermal preferences in office to describe the personalized dynamic thermal comfort in ordinary office
buildings. Building and Environment 118, 323–343. https://doi.org/10.1016/j. environment: from model to application. Building and Environment 72, 309–318.
buildenv.2017.03.009. https://doi.org/10.1016/j.buildenv.2013.11.008.
Li, D., Menassa, C.C., Kamat, V.R., 2017. Personalized human comfort in indoor building Zhou, X., Xu, L., Zhang, J., Niu, B., Luo, M., Zhou, G., Zhang, X., 2020. Data-driven
environments under diverse conditioning modes. Building and Environment 126, thermal comfort model via support vector machine algorithms: Insights from
304–317. https://doi.org/10.1016/j.buildenv.2017.10.004. ASHRAE RP-884 database. Energy and Buildings 211, 109795. https://doi.org/
Lu, S., Wang, W., Lin, C., Hameen, E.C., 2019. Data-driven simulation of a thermal 10.1016/j.enbuild.2020.109795.
comfort-based temperature set-point control with ASHRAE RP884. Building and Zinzi, M., Pagliaro, F., Agnoli, S., Bisegna, F., Iatauro, D., 2021. On the Built-
Environment 156, 137–146. https://doi.org/10.1016/j.buildenv.2019.03.010. Environment quality in nearly Zero-Energy renovated schools: Assessment and
Luo, M., Xie, J., Yan, Y., Ke, Z., Yu, P., Wang, Z., Zhang, J., 2020. Comparing machine impact of passive strategies. Energies 14. https://doi.org/10.3390/en14102799.
learning algorithms in predicting thermal sensation using ASHRAE comfort database Zou, H., Hastie, T., 2005. Regularization and variable selection via the elastic net (vol B
II. Energy and Buildings 210, 109776. https://doi.org/10.1016/j. 67, pg 301. Journal of the Royal Statistical Society, Series B 67 (2005), 768. https://
enbuild.2020.109776. doi.org/10.1111/j.1467-9868.2005.00527.x.

11