International Journal of Heat and Mass Transfer 148 (2020) 119083

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

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

Data-driven prediction of vehicle cabin thermal comfort: using

machine learning and high-fidelity simulation results
Alok Warey a,∗, Shailendra Kaushik a, Bahram Khalighi a, Michael Cruse b,
Ganesh Venkatesan b
a
General Motors Global Research and Development, 30500 Mound Road, Warren, MI 48090, United States
b
Siemens Product Lifecycle Management Software Inc., United States

a r t i c l e i n f o a b s t r a c t

Article history: Predicting thermal comfort in an automotive vehicle cabin’s highly asymmetric and dynamic thermal
Received 15 July 2019 environment is critical for developing energy efficient heating, ventilation and air conditioning (HVAC)
Revised 30 October 2019
systems. In this study we have coupled high-fidelity Computational Fluid Dynamics (CFD) simulations and
Accepted 19 November 2019
machine learning algorithms to predict vehicle occupant thermal comfort for any combination of glazing
Available online 5 December 2019
properties for any window surface, environmental conditions and HVAC settings (flow-rate and discharge
air temperature). A vehicle cabin CFD model, validated against climatic wind tunnel measurements, was
used to systematically generate training data that spanned the entire range of boundary conditions, which
impact occupant thermal comfort. Three machine learning algorithms: linear regression with stochastic
gradient descent, random forests and artificial neural networks (ANN) were applied to the simulation data
to predict the Equivalent Homogeneous Temperature (EHT) for each passenger and the volume averaged
cabin air temperature. The trained machine learning models were tested on unseen data also generated
by the CFD model. Our best machine learning model was able to achieve a test error of less than 5%
in predicting EHT and cabin air temperature. Predicted EHT can also yield thermal comfort metrics such
as Predicted Mean Vote (PMV) and Predicted Percentage of Dissatisfied (PPD), which can account for
different passenger profiles (metabolic rates and clothing levels). Machine learning models developed in
this work enable predictions of thermal comfort for any combination of boundary conditions in real-time
without having to rely on computationally expensive CFD simulations.
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction temperature (typically not equal to air temperature), air velocity

in the vicinity of occupants, humidity and direct solar load on the
Energy efficient heating, ventilation and air conditioning occupants play an equally important role in determination of oc-
(HVAC) systems are critical for maintaining or improving the driv- cupant thermal comfort [2–9].
ing range of electric vehicles (EV’s). A recent study found that on Predicting thermal sensation and comfort in an automobile’s
average, HVAC use at ambient temperature of -7 °C resulted in a thermal environment is both a challenge and an opportunity. It is
41% decrease of combined (city + highway) driving range and a a challenge because of the complex heat transfer between the en-
17% decrease at an ambient temperature of 35 °C (when compared vironment and the human body, and within the body itself, in or-
to testing conducted at 24 °C ambient temperature) [1]. der to maintain thermal homeostasis (primarily through metabolic
The main objective of an automotive HVAC system is to provide processes). On the other hand, it can also be perceived as an op-
thermal comfort to its occupants under various vehicle operating portunity, since an accurate and robust prediction of thermal sen-
conditions. In the highly asymmetric and transient environment of sation and comfort in any arbitrary thermal environment may help
an automobile passenger cabin, the air temperature is only one of HVAC engineers to design and develop a system with optimal per-
many key environmental factors that affect thermal comfort. In ad- formance, low cost and higher energy efficiency [2–9].
dition to the air temperature, the surrounding cabin wall interior Kaushik et al. and Chen et al. developed a CFD model of a
virtual thermal manikin in a passenger cabin. The model was ca-
pable of providing thermal sensations, as experienced by humans,
∗
Corresponding author. under any arbitrary thermal environment. Additionally, the model
E-mail address: [email protected] (A. Warey). could provide sensitivity of thermal sensation to various critical

https://doi.org/10.1016/j.ijheatmasstransfer.2019.119083
0017-9310/© 2019 Elsevier Ltd. All rights reserved.
2 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083

any window surface, environmental conditions and HVAC settings

Nomenclature (flow-rate and discharge air temperature). The methodology en-
ables real-time predictions of thermal comfort for any combination
A/C air conditioning of boundary conditions without having to rely on computationally
ANN artificial neural network expensive CFD simulations.
CFD computational fluid dynamics
CLO clothing insulation, W/mK 1.1. Equivalent homogeneous temperature (EHT)
cp specific heat of air, J/kgK
CV cross-validation It is important to understand that the occupants do not “feel”
DOE design of experiments the cabin air temperature, but they sense the heat loss from the
EHT equivalent homogeneous temperature, °C body. The factors that affect thermal comfort are those that affect
ELU exponential linear unit the body heat loss. Equivalent Homogeneous Temperature (EHT) is
EV electric vehicle a recognized measure of the total heat loss from the human body
HVAC heating, ventilation and air conditioning that can be used to characterize highly non-uniform thermal envi-
IR infrared ronments. It is particularly useful in the confined space of a vehi-
m number of samples cle passenger compartment due to the complex interaction of ra-
m˙ mass flow rate of conditioned air, kg/s diation and convection heat fluxes. The advantage of EHT is that
MAE mean absolute error it expresses the effects of combined thermal influences in a single
MAPE mean absolute percentage error, % variable that is easy to interpret and explain in relation to occu-
MET metabolic activity pant thermal comfort [2,3].
MSE mean squared error EHT is defined as: “The uniform temperature of an imaginary
n_features number of input features enclosure with air velocity equal to zero in which a person will
PMV predicted mean vote exchange the same dry heat by radiation and convection as in the
PPD predicted percentage of dissatisfied, % actual non-uniform environment”, as shown in Fig. 1. The defini-
PRV pressure relief valve tion is based on:
QSolar net solar radiation into the vehicle cabin, W/m2
• A human body exposed to two different environments, one
ReLU rectified linear units
with non-uniform climatic conditions and one with uniform cli-
RMSE root mean square error
matic conditions. It is assumed that the posture, activity level
SGD stochastic gradient descent
and the clothing of the human body are the same in both en-
Tair volume averaged cabin air temperature, °C
vironments.
Tambient ambient temperature, °C
• The dry heat exchange from the human body is the same in
Toutlet HVAC discharge air temperature, °C
both environments.
Twall cabin wall temperature, °C
UA overall thermal conductance of the vehicle cabin, Imagine that we take a person and move him/her from a real
W/K cabin environment into an imaginary room. Now, adjust the tem-
Vair air velocity, m/s perature in the imaginary room until the person experiences the
yi actual or observed value same heat loss as in the real cabin environment. The air tempera-
yˆi predicted value ture in the imaginary room is the Equivalent Homogeneous Tem-
perature. EHT is a pure physical quantity that integrates the inde-
Greek symbols
pendent effects of air temperature, air velocity, mean radiation and
α absorptivity of the window glass
solar load on human body heat loss. This relationship of EHT to the
τ transmissivity of the window glass
surrounding thermal environment best describes the overall whole
body thermal sensation [2,3].
parameters, such as solar load, air conditioning (A/C) discharge
outlet air flow-rate and temperature, A/C outlet locations, various 1.2. Predicted mean vote (PMV)
HVAC modes etc. Comparison of the model’s simulated results
with the actual test data showed good correlation [7,9]. Human perception of thermal comfort is complex and involves
Hintea et al. investigated several machine learning methods both the physiological and the psychological states of a person un-
with the purpose of estimating cabin occupant equivalent tem- der specific conditions. A well-known variable for estimating global
perature from sensors throughout the cabin. Experimental equiva- thermal comfort of people is the Predicted Mean Vote (PMV) index
lent temperature and cabin data at 25 points was gathered from a [11–13]. PMV is used to express the human perception of thermal
test vehicle under different environmental conditions. Seven differ- sensation and is the most widely used thermal comfort index. The
ent machine learning approaches were implemented and evaluated PMV index includes the factors related with thermal comfort, such
[10]. as metabolic activity (MET), clothing insulation (CLO), air tempera-
The present work differs from previous studies in several re- ture, mean radiant temperature, air velocity and humidity. The in-
spects. In this study we have used a vehicle cabin CFD model, vali- dex has been defined by statistical research of a large group of
dated against experimental measurements, to systematically gener- individuals. Zero PMV is the ideal value that represents thermal
ate a comprehensive dataset that spans the entire range of bound- neutrality, and the recommended acceptable PMV range for ther-
ary conditions, which influence occupant thermal comfort. For ex- mal comfort is between -1.0 and +1.0. PMV sensation scale has 7
ample, the generated data includes variation in ambient tempera- levels between -3 and +3 as shown in Table 1 [11–13].
tures from -65 °C to +65 °C. Environmental conditions such as sun Human body heat balance and EHT are dependent on the ther-
position, vehicle orientation, vehicle speed and glazing properties mal resistance of the clothing levels.
in the visible and infrared (IR) spectra for every window surface Clothing level affects the heat loss from the human body and
in the vehicle are included. Machine learning algorithms were ap- plays an important role in determining human thermal sensation.
plied to the simulation data to predict occupant thermal comfort Clothing insulation is expressed in CLO units.1 CLO is defined as
(both front and rear) for any combination of glazing properties for 0.155 W/mK of thermal conductivity [9].
 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083 3

Fig. 1. Schematics for total body heat balance in a vehicle cabin and a uniform environment.

Table 1 Table 2
PMV Sensation Scale [13]. Air flow distribution.

PMV Scale Thermal Sensation Vent % Air Flow

-3 Cold Summer
-2 Cool Center 40
-1 Slightly Cool Left 20
0 Neutral Right 20
1 Slightly Warm Console 20
2 Warm Winter
3 Hot Defrost 15
Front left heater 30
Front right heater 30
Rear left heater 12.5
Rear right heater 12.5

2.1. Vehicle cabin CFD model

In the current study, a commercial CFD package (STAR-CCM+)

was used to model heat transfer between the ambient and the pas-
senger compartment and between the passenger compartment and
the thermal manikins. The vehicle cabin model, as shown in Fig. 3,
was setup in STAR-CCM+ version 13.06.010. A closed volume of the
passenger compartment with a manikin in each seat was meshed
with 16.0 million cells.
The model accounted for all three modes of heat transfer: con-
duction, convection and radiation. It included the conductive heat
Fig. 2. PMV-PPD correlation. transfer between the manikins and the seat, the convective heat
transfer between the conditioned air introduced into the passenger
compartment from the HVAC system and the manikins and finally
1.3. Predicted percentage of dissatisfied (PPD)
radiation heat transfer (both short wave and long wave) radiation
between the passenger compartment’s glazing surfaces and other
PPD is the predicted percentage of dissatisfied people at each
surrounding walls and the manikins. All simulations in this study
PMV and is calculated directly from PMV as follows [12]:
were under steady-state conditions.
P P D = 100 − 95e[−0.3353PMV −0.2179P MV 2 ]
4
(1) The air flow configuration for the air conditioning system un-
der summer conditions is shown in Fig. 4. Air flow distribution for
As PMV deviates from a neutral sensation, PPD increases as both summer and winter conditions is given in Table 2. Summer
shown in Fig. 2. This correlation was based on human subject test conditions were defined as cases where the ambient temperature
data, where the thermal conditions were precisely controlled. 5% was greater than 20 °C. Flow at each vent was directed toward the
PPD is the lowest percentage of dissatisfied practically achievable, passengers’ faces, except the ‘Center’ vent, which was split to be
since providing an optimal thermal environment for every single directed at the front passengers’ chests as shown in Fig. 5.
person is not possible. The recommended acceptable PPD range for The air flow configuration for the heating system under winter
the thermal comfort is less than 10% PPD [12]. conditions is shown in Fig. 6. Winter conditions were defined as
cases where the ambient temperature was less than 20 °C.
2. Data generation The conditioned air entering the passenger compartment leaves
via PRVs (pressure relief valves). In the model, the PRVs are mod-
The data used for training the machine learning models in this eled as pressure outlets. Their locations have an important impact
work was obtained from high-fidelity CFD simulations of the vehi- on the flow-field pattern in the compartment and hence on the
cle cabin. convective heat transfer of the manikins.
4 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083

Fig. 3. Physical configuration simulated in the model.

Fig. 4. Air conditioning system air flow configuration.

Appropriate thermo-physical properties, were imposed on each prescribed as heat transfer coefficients and ambient air tempera-
wall. In addition, thermal radiation properties such as emissiv- ture, reflecting the speed of the vehicle.
ity, reflectivity and transmissivity for each wall were prescribed. The manikin’s surfaces were segmented into 16 body parts.
Since solar radiation plays a significant role in the overall heat Each of these parts was covered with clothing of constant insu-
balance inside the passenger compartment, glazing properties for lation, specified in units of CLO. Light summer clothing, for ex-
each window surface were considered in the data generation. ample, had an average overall insulation value of 0.6 CLO. Higher
An external convective boundary condition for these walls, was level of clothing provides less sensitivity to surrounding thermal
A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083 5

Fig. 5. Air flow velocity distribution for the A/C system.

Fig. 6. Heating system air flow configuration.

Fig. 7. Basic energy balance of a passenger compartment at steady-state.

6 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083

Table 3
Climatic wind tunnel boundary conditions.

Solar Load (W/m2 ) Total HVAC Flow Rate (CFM) Discharge Air Temperature (°C) Ambient Temperature (°C) Vehicle Speed (kph)

0 100 50 -20 112

tunnel test. The volume averaged cabin air temperature from the
CFD model matched the target cabin air temperature of 24 °C as
measured in the tunnel test, which resulted in a UA value of 35
W/mK (same as wind tunnel data) as shown in Fig. 8. Overall EHT
values for the ‘virtual’ thermal manikins were validated with those
obtained from a physical manikin in the climatic wind tunnel tests
under steady-state boundary conditions.
Additional validation under typical summer conditions was car-
ried out with data from a test vehicle. Under the same boundary
conditions, the volume averaged cabin air temperature calculated
in the CFD model matched experimental measurements within
±1 °C.

2.3. Simulation dataset

A design of experiments (DOE) matrix of glazing properties in

the visible and infrared (IR) spectra for every window surface in
Fig. 8. Overall thermal conductance with vehicle speed measured in a climatic the vehicle, environmental boundary conditions and HVAC settings
wind tunnel at -20 °C ambient temperature for 24 °C cabin air temperature. (discharge air temperature and flow-rate) for a total of 37 features
was generated as shown in Table 4. The glazing properties (absorp-
tivity and transmissivity) in the visible and IR spectra were con-
environment due to high thermal resistance. To compute the net
strained in the DOE matrix to account for reflectivity of the win-
heat flux at a body segment, a constant skin temperature of 34°C
dow surface as follows:
was used for each passenger manikin. Occupant heat flux data
from the CFD simulations was used to calculate an overall EHT for α+ τ ≤1 (3)
each passenger as outlined by Han et al. and Kaushik et al. [3,7].
where α is the absorptivity and τ is the transmissivity of a win-
dow surface in the visible or IR spectra. Since reflectivity is a de-
2.2. Vehicle cabin model validation pendent variable on absorptivity and transmissivity, it was left out
of the DOE matrix.
The vehicle cabin CFD model was validated with climatic wind CFD simulations were performed at each design point in the
tunnel measurements. The basic energy balance of the passenger DOE matrix. Occupant heat flux data from the CFD simulations
compartment at a steady-state condition is given by: was used to calculate an overall EHT for each passenger for all
m˙ c p (Toutlet − Tair ) = UA(Tair − Tambient ) − QSolar (2) cases in the DOE matrix. 241 cases were run to generate the train-
ing dataset. An additional 60 cases that covered typical summer
where m˙ is the mass flow rate of conditioned air into the cabin, and winter conditions and varying glazing properties for the sun-
cp is the specific heat of air, Toutlet is the HVAC discharge air tem- roof were run to generate the test dataset. Machine learning model
perature, Tair is the cabin air temperature, Tambient is the ambient training and validation was carried out with the training dataset.
temperature, QSolar is the net solar radiation into the cabin and UA The test dataset was only used to check the performance of the
is the overall thermal conductance of the cabin enclosure. optimized machine learning models on unseen data.
The right side of Eq. (2) represents the environmental thermal
load in the passenger compartment on the HVAC system, while 3. Machine learning models
the left side terms represent the cooling or heating power pro-
vided by the HVAC system. As shown in Fig. 7, for a given ambient Three machine learning algorithms, a linear model, an ensem-
temperature and solar load, correct amount of HVAC power must ble predictor and an artificial neural network, were investigated to
be provided to maintain a desired cabin air temperature for occu- predict the overall EHT for each passenger and volume averaged
pant thermal comfort. The environmental thermal load in the cabin cabin air temperature from the input training data described ear-
(second term in Eq. (2)) has two terms. The first term is the overall lier.
heat exchange between the cabin and the environment, which is
proportional to the temperature difference between the cabin inte- 3.1. Linear regression with stochastic gradient descent
rior and the ambient air and the thermal conductance of the cabin
enclosure, UA. The second term is the net solar radiation into the A linear regression model using stochastic gradient descent [14–
cabin. 16] was implemented with the machine learning library Scikit-
The overall thermal conductance, UA, of the passenger cabin learn in Python [17]. A multi-target regression strategy of fitting
is calculated from the climatic wind tunnel data as a function of one regressor per target was used. This is a simple strategy for
vehicle speed at a very cold ambient temperature (for example, extending regressors that do not natively support multi-target re-
-20 °C) without solar load. The overall thermal conductance of gression. All input features were zero centered and normalized
the cabin depends mainly on the vehicle architecture and vehicle by the mean and standard deviation of the training dataset. The
speed. Overall vehicle thermal conductance measured from a hyperparameters were tuned by conducting a randomized search
climatic wind tunnel test is shown in Fig. 8. The CFD model was over the hyperparameter space shown in Table 5 with 6-fold cross-
run at the same boundary conditions given in Table 3 as the wind validation (CV) over the training dataset. The final hyperparameter
 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083 7

Table 4
Features included in the training data.

Feature Name Min. Max. Units

Vehicle Orientation 0 360 degrees

Solar Load 0 1000 W/m2
Sun Altitude 0 90 degrees
Sun Azimuth 0 360 degrees
Flow Rate 50 300 CFM
Discharge Air Temperature 4 65 °C
Ambient Temperature -65 65 °C
Clothing Level 0.3 2 CLO units
Vehicle Speed 0 70 mph
Sunroof Absorptivity (Visible) 0 1 [-]
Sunroof Transmissivity (Visible) 0 1 [-]
Sunroof Absorptivity (Infrared) 0 1 [-]
Sunroof Transmissivity (Infrared) 0 1 [-]
Windshield Absorptivity (Visible) 0 1 [-]
Windshield Transmissivity (Visible) 0 1 [-]
Windshield Absorptivity (Infrared) 0 1 [-]
Windshield Transmissivity (Infrared) 0 1 [-]
Left Front Side Window Absorptivity (Visible) 0 1 [-]
Left Front Side Window Transmissivity (Visible) 0 1 [-]
Left Front Side Window Absorptivity (Infrared) 0 1 [-]
Left Front Side Window Transmissivity (Infrared) 0 1 [-]
Right Front Side Window Absorptivity (Visible) 0 1 [-]
Right Front Side Window Transmissivity (Visible) 0 1 [-]
Right Front Side Window Absorptivity (Infrared) 0 1 [-]
Right Front Side Window Transmissivity (Infrared) 0 1 [-]
Left Rear Side Window Absorptivity (Visible) 0 1 [-]
Left Rear Side Window Transmissivity (Visible) 0 1 [-]
Left Rear Side Window Absorptivity (Infrared) 0 1 [-]
Left Rear Side Window Transmissivity (Infrared) 0 1 [-]
Right Rear Side Window Absorptivity (Visible) 0 1 [-]
Right Rear Side Window Transmissivity (Visible) 0 1 [-]
Right Rear Side Window Absorptivity (Infrared) 0 1 [-]
Right Rear Side Window Transmissivity (Infrared) 0 1 [-]
Rear Window Absorptivity (Visible) 0 1 [-]
Rear Window Transmissivity (Visible) 0 1 [-]
Rear Window Absorptivity (Infrared) 0 1 [-]
Rear Window Transmissivity (Infrared) 0 1 [-]

Table 5
Hyperparameter optimization of the linear regression model.

Parameter Search Space Final Values

Learning rate [0.001, 0.01, 0.1] 0.01

Regularizer [18] [L2 norm, L1 norm] L2 norm
Constant that multiplies the regularization term [0.00001, 0.0001, 0.001] 0.0001
Number of epochs [500, 1000, 1500] 1000

Table 6
Hyperparameter optimization of the random forest model

Parameter Search Space Final Values

Number of estimators [10, 50, 100, 500, 1000] 100

Maximum depth of the trees [5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, None] None
Number of features to consider when looking for the best split [n_features, sqrt(n_features)] n_features
Maximum leaf nodes [5, 10, 15, 20, None] None

values chosen for the linear regression model are also shown in of the trees and maximum leaf nodes for the trees were tuned
Table 5. by conducting a randomized search over the hyperparameter space
shown in Table 6 with 6-fold cross-validation (CV) over the train-
3.2. Random forests ing dataset. The final hyperparameter values chosen for the ran-
dom forest model are also shown in Table 6.
Random forests are an ensemble learning method that operates
by fitting a number of decision trees on various sub-samples of the 3.3. Artificial neural network
training dataset [19]. It was implemented with the machine learn-
ing library Scikit-learn in Python [17]. Although random forests A fully connected feedforward neural network as shown in
natively support multi-output regression, a multi-target regression Fig. 9 was developed to predict the EHT for each passenger and
strategy of fitting one random forest per target was used. This volume averaged cabin air temperature. In a feedforward network,
strategy resulted in better performance during cross-validation. the first layer is the input and the last layer is the output, in be-
The hyperparameters such as number of estimators, maximum tween are the hidden layers. The units in hidden layer are termed
number of features to consider for the best split, maximum depth “neurons”. The ANN can be regarded as a process where the input
8 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083

Fig. 9. Artificial neural network.

Table 7
ANN architecture and hyperparameter optimization.

Parameter Search Space Final Values

Neurons in the first hidden layer [8, 16, 32, 64, 128, 256] 256
Additional hidden layers [0, 1, 2] 1
Neurons in the additional hidden layers [8, 16, 32, 64, 128, 256] 64
Activation [25,26] [ReLU, ELU] ELU
Optimizers [16] [Adam, RMSprop, SGD] Adam
Dropout after each hidden layer [27] [0, 0.1, 0.2, 0.3, 0.4, 0.5] 0.2
Batch size [1, 4, 8, 16, 32] 8
Learning rate [0.001, 0.01, 0.1] 0.001

Table 8
Summary statistics of machine learning models over 10 cross-validation folds.

EHT: Front Left EHT: Front Right EHT: Rear Left EHT: Rear Right Cabin Air Temperature

Linear Regression with Stochastic Gradient Descent

Average MAE [°C] 4.5 4.8 4.8 4.7 4.9
Std. Deviation [°C] 0.8 1.0 0.8 0.7 1.0
Random Forest
Average MAE [°C] 3.2 3.7 3.8 4.0 3.7
Std. Deviation [°C] 0.4 0.7 0.5 0.7 0.3
Artificial Neural Network
Average MAE [°C] 2.1 2.5 2.3 2.5 2.6
Std. Deviation [°C] 0.4 0.4 0.4 0.3 0.4

features go through a series of nonlinear transformations to predict where m is the number of samples, yˆi is the predicted value and
the outputs [20]. In this work, the open source deep learning li- yi is the actual or observed value.
brary Keras [21,22] with TensorFlow [23] as the backend was used
for constructing, training and applying the ANNs. All input features 3.4. Model evaluation
were zero centered and normalized by the mean and standard de-
viation of the training dataset. The machine learning models were evaluated on the full set of
The neural network architecture and hyperparameters were op- training data using 10-fold cross- validation (CV), indicating how
timized using Talos the open source hyperparameter optimiza- well the network generalized to unseen data. Table 8 shows the
tion solution for Keras [24]. 20% of the training data was used summary statistics for the three models over the 10 folds. Mean
for validation during optimization. The network architecture pa- absolute error for each fold was defined as:
1   
rameters considered in the optimization are shown in Table 7. m
Over 50 0 0 network hyperparameter combinations were evaluated Mean Absolute Error (MAE ) = yˆi − yi  (5)
by conducting a randomized search through the entire parameter
m
i=1
space over several iterations. At each iteration the parameter space Even with hyperparameter tuning, linear regression with
was shrunk based on results from the previous iteration. The final stochastic gradient descent performed the worst with an average
network configuration also shown in Table 7 was chosen based on mean absolute error over the 10 folds of approximately 4.5 °C.
the loss (Mean Squared Error) on the validation dataset as given Overall, the optimized ANN performed the best among all three
by: machine learning models with an average mean absolute error
1  2
m over the 10 folds of approximately 2.5 °C.
Mean Squared Error (MSE ) = yˆi − yi (4) P-values were generated through Wilcoxon signed-rank test
m
i=1 for the ANN vs. linear regression with SGD and random forests to
 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083 9

Fig. 10. Predicted EHT values (°C) by the ANN vs. EHT values from CFD simulations for all four passengers over the test set of 60 cases.

Table 9 establish the significance of these results as given in Table 9. Low

P-values for ANN vs. linear regression with SGD and random forests.
p-values indicate that the ANN was significantly better than linear
Model 1 Model 2 p-value regression with stochastic gradient descent and random forests.
EHT: Front Left Linear Regression with SGD ANN 5.1E-03
Random Forests 6.9E-03
EHT: Front Right Linear Regression with SGD ANN 6.9E-03
Random Forests 1.7E-02
EHT: Rear Left Linear Regression with SGD ANN 5.1E-03 4. Results
Random Forests 5.1E-03
EHT: Rear Right Linear Regression with SGD ANN 6.9E-03
4.1. Performance on the test dataset
Random Forests 1.7E-02
Cabin Air Temperature Linear Regression with SGD ANN 6.9E-03
Random Forests 1.3E-02 The optimized ANN was used to make predictions on a test
set of 60 samples that were held out as unseen data. The test set
spanned both summer and winter conditions. In addition to the
mean absolute error, following metrics were used to evaluate the

Table 10
Performance metrics of the ANN on test data.

EHT: Front Left EHT: Front Right EHT: Rear Left EHT: Rear Right Cabin Air Temperature

MAE [°C] 0.8 0.8 0.8 0.7 1.1

RMSE [°C] 1.0 0.9 0.9 0.9 1.4
MAPE [%] 3.3 2.9 3.3 3.0 4.7
10 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083

Table 11
Sunroof glass properties.

Absorptivity (Visible) Transmissivity (Visible) Absorptivity (Infrared) Transmissivity (Infrared)

[-] [-] [-] [-]

Sunroof Glass I
0.22 0.72 0.72 0.22
Sunroof Glass II
0.93 0.02 0.93 0.03

Table 12
Summer boundary conditions.

Solar Load (W/m2 ) Total HVAC Flow Rate (CFM) Ambient Temperature (°C) Vehicle Speed (mph)

1000 200 30 70

Fig. 11. PMV-PPD for all four passengers for the two sunroof glasses under summer boundary conditions.

Table 13 performance of the ANN on unseen data:

Occupant biometrics. 
1  2
m
Weight (kg) 57
Height (cm) 165
Root Mean Square Error (RMSE ) = yˆi − yi (6)
m
Age (years) 40 i=1
Sex Female
 
100%   yˆi − yi 
m
Mean Absolute Percentage Error (MAP E ) =  yi 
m
i=1
(7)
 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083 11

Table 14
HVAC settings for the two sunroof glasses.

Flow Rate Discharge Air Temperature Predicted Cabin Air Temperature Calculated HVAC Power
CFM °C °C W

Sunroof Glass I
200 4 18.6 2463
Sunroof Glass II
200 9 21.3 1989

Performance metrics of the ANN on unseen data are given networks were applied to the simulation data to predict the equiv-
in Table 10. After training on the entire training dataset, perfor- alent homogeneous temperature (EHT) for each passenger and the
mance of the ANN on the test set improved compared to the volume averaged cabin air temperature.
cross-validation datasets. The optimized ANN was able to achieve The machine learning models were evaluated on the full set of
a RMSE of less than 1.5 °C and a mean absolute percentage error training data using 10-fold cross- validation (CV), indicating how
of less than 5% on the test data. well the network generalized to unseen data. Wilcoxon signed-
Fig. 10 shows the comparison between the predicted EHT values rank test for the ANN vs. linear regression with SGD and random
(°C) by the optimized ANN and EHT values from CFD simulations forests indicated that the ANN was significantly better. The opti-
for all four passengers over the test set of 60 cases. mized ANN was able to achieve a RMSE of less than 1.5 °C and a
mean absolute percentage error of less than 5% on the test data.
The optimized ANN was used to assess the impact of sunroof
4.2. Case study: impact of glazing on HVAC energy consumption
glazing on energy consumption of the HVAC system under typical
summer conditions. The case study highlights the promise of using
The optimized ANN was used to assess the impact of sunroof
data-driven machine learning models to predict thermal comfort
glazing on energy consumption of the HVAC system under steady-
for any combination of boundary conditions in real-time without
state boundary conditions. Glazing properties of the two sunroof
having to rely on computationally expensive CFD simulations.
glasses considered in the study are given in Table 11. Glazing prop-
erties for all other window surfaces in the vehicle were held con- Declaration of Competing Interest
stant.
The two glasses were evaluated under typical summer bound- None.
ary conditions given in Table 12 with the sun directly overhead
(Sun altitude = 90° and Sun Azimuth = 0°) Supplementary materials
All four occupants were assumed to have identical biometrics
given in Table 13 below. Since PMV-PPD calculations are done us- Supplementary material associated with this article can be
ing predictions of EHT, these traits can be easily modified to ac- found, in the online version, at doi:10.1016/j.ijheatmasstransfer.
commodate different passenger profiles. 2019.119083.
To maintain the same level of thermal comfort for all four pas-
sengers as shown in Fig. 11, the HVAC settings (flow rate and References
discharge air temperature) for the two glass types are given in
[1] AAA Newsroom, Icy Temperatures Cut Electric Vehicle Range
Table 14 along with the predicted cabin air temperature and HVAC Nearly in Half, https://www.aaa.com/AAA/common/AAR/files/
power calculated as: AAA- Electric- Vehicle- Range- Testing- Report.pdf, 2019 (accessed 18 June
2019).
HVAC Power = m˙ c p (Tambient − Toutlet ) (8) [2] T. Han, L. Huang, S. Kelly, C. Huizenga, Z. Hui, Virtual Thermal Comfort Engi-
neering, SAE Technical Paper 2001-01-0588, 2001, doi:10.4271/2001-01-0588.
[3] T. Han, L. Huang, A Model for Relating a Thermal Comfort Scale to EHT Comfort
where m˙ is the mass flow rate of conditioned air, cp is the specific
Index, SAE Technical Paper 2004-01-0919, 2004, doi:10.4271/2004- 01- 0919.
heat of air, which was assumed to be 1003.62 J/kgK, Tambient is the [4] T. Han, L. Huang, A Sensitivity Study of Occupant Thermal Comfort in a Cabin
ambient air temperature and Toutlet is the discharge air tempera- Using Virtual Thermal Comfort Engineering, SAE Technical Paper 2005-01-
1509, 2005, doi:10.4271/2005- 01- 1509.
ture into the cabin. To maintain the same level of thermal comfort
[5] T. Han, K. Chen, Assessment of Various Environmental Thermal Loads on Pas-
for all four passengers under identical summer boundary condi- senger Compartment Soak and Cool-down Analyses, SAE Technical Paper 2009-
tions, the discharge air temperature for the low transmissivity glass 01-1148, 2009, 10.4271/2009-01-1148.
(Sunroof Glass II) could be set at 9 °C vs. 4 °C for the clear glass [6] S. Kaushik, K. Chen, T. Han, B. Khalighi, Micro-cooling/heating strategy for en-
ergy efficient HVAC system, SAE Int. J. Mater. Manuf. 4 (1) (2011) 853–863,
(Sunroof Glass I), which resulted in a HVAC power consumption doi:10.4271/2011-01-0644.
reduction of 474 W. [7] S. Kaushik, T. Han, K. Chen, Development of a Virtual Thermal Manikin to
The case study highlights the promise of using data-driven ma- Predict Thermal Sensation in Automobiles, SAE Technical Paper 2012-01-0315,
2012, doi:10.4271/2012-01-0315.
chine learning models to do preliminary design studies of HVAC [8] D. Ghosh, M. Wang, E. Wolfe, K. Chen, S. Kaushik, T. Han, Energy efficient HVAC
systems in real-time without having to rely on computationally ex- system with spot cooling in an automobile - design and CFD analysis, SAE Int.
pensive CFD simulations or wind tunnel tests. J. Passeng. Cars Mech. Syst. 5 (2) (2012) 885–903, doi:10.4271/2012- 01- 0641.
[9] K. Chen, S. Kaushik, T. Han, D. Ghosh, M. Wang, Thermal Comfort Prediction
and Validation in a Realistic Vehicle Thermal Environment, SAE Technical Pa-
per 2012-01-0645, 2012, doi:10.4271/2012-01-0645.
5. Summary
[10] D. Hintea, J. Brusey, E. Gaura, A study on several machine learning methods
for estimating cabin occupant equivalent temperature, in: Proceedings of the
High-fidelity CFD simulations and machine learning algorithms Twelfth International Conference on Informatics in Control, Automation and
Robotics (ICINCO), Colmar, France, 2015, doi:10.4271/2012-01-0645.
were coupled to develop a data-driven tool to predict occupant
[11] P.O. Fanger, Thermal Comfort: Analysis and Applications in Environmental En-
thermal comfort for any combination of glazing properties for gineering, Danish Technical Press, Copenhagen, 1970.
any window surface, environmental conditions and HVAC settings [12] N. Djongyang, R. Tchinda, D. Njomo, Thermal comfort: a review paper, Renew.
(flow-rate and discharge air temperature). Sustain. Energy Rev. 14 (9) (2010) 2626–2640, doi:10.1016/j.rser.2010.07.040.
[13] S. Gilani, M.H. Khan, W. Pao, Thermal comfort analysis of PMV model pre-
Three machine learning algorithms: linear regression with diction in air conditioned and naturally ventilated buildings, Energy Proc. 75
stochastic gradient descent, random forests and artificial neural (2015) 1373–1379, doi:10.1016/j.egypro.2015.07.218.
12 A. Warey, S. Kaushik and B. Khalighi et al. / International Journal of Heat and Mass Transfer 148 (2020) 119083

[14] J. Kiefer, J. Wolfowitz, Stochastic estimation of the maximum of a regression [22] F. Chollet, Deep Learning with Python, 1st ed., Manning Publications Co., Shel-
function, Ann. Math. Stat. 23 (3) (1952) 462–466. ter Island, 2018.
[15] L. Bottou, F.E. Curtis, J. Nocedal, Optimization Methods for Large-Scale Machine [23] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat,
Learning, 2018, arXiv preprint arXiv:1606.04838. G. Irving, M. Isard, M. Kudlur, J. Levenberg, R. Monga, S. Moore, D. G. Murray, B.
[16] S. Ruder, An Overview of Gradient Descent Optimization Algorithms, 2017, Steiner, P. Tucker, V. Vasudevan, P. Warden, M. Wicke, Y. Yu, X. Zheng, Tensor-
arXiv preprint arXiv:1609.04747. Flow: a system for large-scale machine learning, in: Proceedings of the Twelfth
[17] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, USENIX Symposium on Operating Systems Design and Implementation, Savan-
M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, nah, USA, 2016.
D. Cournapeau, M. Brucher, M. Perrot, É. Duchesnay, Scikit-learn: machine [24] M. Kotila, Talos, (2018), GitHub repository, https://github.com/autonomio/talos
learning in python, J. Mach. Learn. Res. 12 (2011) 2825–2830. [25] V. Nair, G.E. Hinton, Rectified linear units improve restricted Boltzmann ma-
[18] A. Ng, Feature Selection, L1 vs. L2 Regularization, and Rotational Invariance, in: chines, in: Proceedings of the Twenty-Seventh International Conference on
Proceedings of the Twenty-First International Conference on Machine Learning, Machine Learning, Haifa, Israel, 2010.
Banff, Canada, 2004, doi:10.1145/1015330.1015435. [26] D. Clevert, T. Unterthiner, S. Hochreiter, Fast and accurate deep network learn-
[19] L. Breiman, Random Forests, Machine Learning 45 (1) (2001) 5–32, doi:10. ing by exponential linear units (ELUs), in: Proceedings of the Fourth Interna-
1023/A:1010933404324. tional Conference on Learning Representations, San Juan, Puerto Rico, 2016.
[20] A. Geron, Hands-On Machine Learning with Scikit-Learn and TensorFlow, 1st [27] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout:
ed., O’Reilly Media Inc., Sebastopol, 2017. a simple way to prevent neural networks from overfitting, J. Mach. Learn. Res.
[21] F. Chollet, Keras, (2015), GitHub repository, https://github.com/keras-team/ 15 (2014) 1929–1958.
keras