Economic and Efficient Hybrid Vehicle Fuel Economy and Emissions Modeling Using an Artificial Neural Network

Abstract

High accuracy hybrid vehicle fuel consumption (FC)
and emissions models used in practice today are the
product of years of research, are physics based, and
bear a large computational cost. However, it may be possible
to replace these models with a non-physics based, higher
accuracy, and computationally efficient versions. In this
research, an alternative method is developed by training and
testing a time series artificial neural network (ANN) using
real world, on-road data for a hydraulic hybrid truck to predict
instantaneous FC and emissions. Parameters affecting model
fidelity were investigated including the number of neurons in
the hidden layer, specific training inputs, dataset length, and
hybrid system status. The results show that the ANN model
was computationally faster and predicted FC within a mean
absolute error of 0-0.1%. For emissions prediction the ANN
model had a mean absolute error of 0-3% across CO2, CO,
and NOx aggregate predicted concentrations. Overall, these
results indicate that ANN models could be used for a variety
of research applications due to their economic and computational
benefits such as derivation of vehicle control strategies
to reduce FC and emissions in modern vehicles.

Full text

2018-01-0315 Published 03 Apr 2018
© 2018 SAE International. All Rights Reserved.
Economic and Ecient Hybrid Vehicle Fuel
Economy and Emissions Modeling Using
anArtificial Neural Network
Zachary D. Asher and Abril A. Galang Colorado State University
Will Briggs and Brian Johnston Lightning Systems
Thomas H. Bradley and Shantanu Jathar Colorado State University
Citation: Asher, Z.D., Galang, A.A., Briggs, W., Johnston, B. et al., “Economic and Efficient Hybrid Vehicle Fuel Economy and Emissions
Modeling Using an Artificial Neural Network,” SAE Technical Paper 2018-01-0315, 2018, doi:10.4271/2018-01-0315.
Abstract
High accuracy hybrid vehicle fuel consumption (FC)
and emissions models used in practice today are the
product of years of research, are physics based, and
bear a large computational cost. However, it may be possible
to replace these models with a non-physics based, higher
accuracy, and computationally efficient versions. In this
research, an alternative method is developed by training and
testing a time series articial neural network (ANN) using
real world, on-road data for a hydraulic hybrid truck to predict
instantaneous FC and emissions. Parameters aecting model
delity were investigated including the number of neurons in
the hidden layer, specic training inputs, dataset length, and
hybrid system status. e results show that the ANN model
was computationally faster and predicted FC within a mean
absolute error of 0-0.1%. For emissions prediction the ANN
model had a mean absolute error of 0-3% across CO2, CO,
and NOx aggregate predicted concentrations. Overall, these
results indicate that ANN models could be used for a variety
of research applications due to their economic and computa-
tional benets such as derivation of vehicle control strategies
to reduce FC and emissions in modern vehicles.
Introduction
The transportation sector accounts for 28% of all energy
use in the U.S. and 30% globally [1, 2]. Most of this
energy is generated from the combustion of fossil fuels,
which produces greenhouse gases (GHG) such as CO2 and
CH4,and harmful pollutants such as CO, NOx, hydrocarbons
(HC), and particulate matter less than 2.5μm in diameter
(PM2.5), that aect air quality, climate and human health [3,
4, 5, 6, 7, 8, 9]. ere is an urgent need to reduce vehicle fuel
consumption (FC) and emissions of harmful pollutants.
Corporate Average Fuel Economy (CAFE) standards enacted
by the US Congress in 1975 must be met by individual auto-
makers for its car and truck eet [10]. Recently, in 2012, CAFE
standards were updated to project a eet wide average FC of
40.3-41.0 mpg by model year 2021. is goal operates in
tandem with the Environmental Protection Agency’s (EPA)
goal to limit CO2 emissions to 163 grams/mile by model year
2025. In addition, state (e.g., California Air Resources Board)
and federal (e.g., Environmental Protection Agency) agencies
regulate other emissions such as CO, NOx, HC, and PM.
A key technology to decrease FC and emissions is vehicle
hybridization and implementation of advanced control strate-
gies [11]. Decreasing FC will likely, but not necessarily, reduce
total emissions produced due to less fuel consumed. However,
the emissions composition may change with vehicle specic
power (VSP) and steady state engine operation [12, 13]. e
EPA has developed the MOtor Vehicle Emission Simulator
(MOV ES) [14, 15, 16, 17, 18] to model emissions from on-road
mobile sources. While not perfect, MOVES oers satisfactory
performance for conventional vehicles but it has been shown
that the model does not perform as well with HEVs [19].
Emissions tests on dynamometers (a device that measures and
imposes torque and power on an engine or vehicle) using
certication cycles tend to produce less pollutants (10-20%)
when compared to real world, on-road test conditions [20].
Tests completed on a chassis dynamometer cannot fully mimic
turns and potholes, unexpected and inopportune trac inci-
dents, and aggressive or uncommon driving habits. ese are
characterized primarily in real world, on-road collected data.
In order to properly evaluate FC and emissions from
hybrid vehicle architecture and alternate control strategies,
extensive and expensive testing or modeling is required.
Vehicle FC and emissions characterization is typically tested
on a chassis dynamometer or from physical vehicle operation
in the real world. Design and control development research
is typically simulated using modeling software such as
Autonomie due to its demonstrated correlation with physical
vehicle operation [21, 22]. However, physics-based simulations
such as Autonomie are computationally costly and typically
provide unneeded vehicle performance results, thus it has
Downloaded from SAE International by Colorado State Univ, Wednesday, April 04, 2018

2 ECONOMIC AND EFFICIENT HYBRID VEHICLE FUEL ECONOMY AND EMISSIONS MODELING
© 2018 SAE International. All Rights Reserved.
become standard practice to derive improved control strate-
gies using simplied, equation-based hybrid vehicle models
[23]. e drawback of using a simplied vehicle model is that
they do not provide high accuracy results [24]. Neither of these
two solutions provides the compromise between accurac y and
speed that is required for vehicle design and control research.
In addition to vehicle testing and modeling consider-
ations, there is also the consideration that modern and near
future vehicles may benet from an internal model of FC and
emissions for daily operation. Numerous vehicle components
incorporate learning controllers to improve performance and
adapt to component wear [25, 26]. As the automotive industry
pushes to implement Eco-Driving and Optimal Energy
Management Strategies [27] a learning controller focused on
overall FC and emissions may be required. But, traditional
learning controllers such as a Kalman lter do not provide
the number of inputs required to capture FC and emissions
and new learning controllers such as ar ticial neural networks
may be benecial. is has led to research eorts investigating
their potential automotive applications.
Articial neura l networks (ANN) are computing systems
inspired by biological neural networks that improve perfor-
mance without task specic progra mming [28]. e rst prac-
tical application of an articial neural network was in the
1950s [29] but did not experience mainstream application
until understanding of statistical mechanics [30] and the
development of back propagation in the 1980s [31]. Since then,
articial neural networks have been a powerful tool when
using a large set of data where the important features are
learned by the algorithms instead of being pre-programmed
by humans [28]. Articial neural networks have successfully
been applied to multiple automotive problems with the most
focus being on predicting engine performance [32, 33, 34, 35,
36, 37]. ese studies focus on engine FC and emissions using
primarily dynamometer data and none have been used to
capture the complex dynamics of hybrid vehicle system FC
and emissions in real world operation. Additionally, none of
these studies consider vehicle learning control implementa-
tion for Eco-Driving and Optimal Energy Management.
In this study, we perform FC and emissions measurements
with a portable emissions monitoring system (PEMS) on a
hybrid vehicle and use that data to develop an ANN model for
that hybrid vehicle. We evaluate the constructed ANN model
against independent data. Our work suggests that the use of
an ANN model, in contrast to a traditional physics-based
model, reduces the computation time but does not sacrice
accuracy. e use of an ANN model, if developed for a vehicle
and integrated within t he vehicle controller, would be extremely
valuable for vehicle/engine development and producing real-
time reductions in fuel consumption and emissions [24].
Methods
We use a neural network framework to develop and train
models to predict the FC and emissions from hybrid vehicles.
e models are trained on data gathered using a portable
emissions monitoring system (PEMS) on a hydraulic
hybrid truck.
Neural Network Models
NNs are trained iteratively to change the weights and biases
in a neuron with the transfer function:
nxwb
ii
i
p
=+
=
å
1
(1)
where n is the neuron output, P is the number of input
elements, xi is the input value, wi is the weight of the input
value, and b is the bias. Weights and biases are altered with
each iteration, inuencing the output relative to the impor-
tance of the input [38]. e summation of the weighted inputs
and the network biases yield an output, n, in the hidden layer
[39]. Figure 1 shows the architecture of a Nonlinear
AutoRegressive network with eXogenous (NARX) inputs,
which is the form of all ANNs in this study. A sigmoid transfer
function is used in the hidden layer neurons producing an
output from 0 to 1. Using a sigmoid function in the hidden
layer allows the weights and biases to generalize an output
unlike a step function. Neurons in the hidden layer pass
predictions to a single neuron in the output layer. e output
layer uses a linear transfer function to scale the overall output
for the target units and scale. Time series datasets utilize more
inputs dependent on a time delay d:
y
tfyt yt yt d
()
=-
()
-
()
¼-
()
[,
,,12
xt
xt xt d
-
()
-
()
¼-
()
12,,
,]
(2)
Where y(t) is the network output, x is the training input,
and t is the time step. e time delay determines the amount
of time each data point is included in the transfer functions.
e performance of ANNs are evaluated by calculating
the correlation coecient (R) and the mean square error
(MSE) in each training iteration:
RapEaupu
Eaupu
ap
ap
,
()
=-
()
-
é
ëù
û
-
()
-
()
(
22 (3)
MSE
npa
ii
i
n
=-
()
=
å
12
1
(4)
where E is the expected value, a and p are measured and
predicted output sets respectively, ua and up are the mean va lues
of the a and p sets respectively, and n is the length of the a and
p sets. e R-value, R(a, p), is a measure of how well the predic-
tions correlate linearly to the measurements. e MSE measures
FIGURE 1 Schematic for a NARX trained to predict FC and
emissions predictions. In this case, 4 inputs (x(t)) and 1 feed-
forward output (y(t)) enter the network. The output of the
network is used as an input from pervious time steps. The time
delay for this network is 3seconds (1:3) and there are 15
neurons in the Hidden Layer [39].
© SAE International
Downloaded from SAE International by Colorado State Univ, Wednesday, April 04, 2018

ECONOMIC AND EFFICIENT HYBRID VEHICLE FUEL ECONOMY AND EMISSIONS MODELING 3
© 2018 SAE International. All Rights Reserved.
average squared error in the prediction. e training process
attempts to maximize the R value and minimize the MSE. e
weights and biases are improved using backpropagation algo-
rithms in an iterative process [38]. Each iteration builds on the
previous network, changing weights and biases, to optimize
MSE and R. In this work, the performance of the trained
networks was judged by calculating the percent error in the
measured and predicted FC over an entire cycle:
%Error
measured predicted
measured
i
i
n
i
i
n
i
i
n
=-*
==
=
åå
å
11
1
100 (5)
Where n is the number of predicted and measured values,
measuredi is the measured value, and predictedi is the model
predicted value. e sign of percent error calculated depicts
whether a model under or over estimates a prediction.
Training networks from comprehensive real world driving
data can create networks that can predict the FC or emissions
for any drive cycle.
Portable Emissions Monitoring
System
Portable Emissions Monitoring System (PEMS) devices have
been used to measure on-road emissions in real world trac
[40, 41, 42, 43, 44, 45, 46, 47, 48] and increasingly oered similar
performance to reference instrumentation used in laboratories
[49]. e emissions data in this study was collected using an
AxionS+ PEMS manufactured by Global MRV. Similar PEMS
devices from the same manufacturer have performed well and
have been used on many dierent vehicles including light duty
vehicles, refuse trucks, and conventional and hybrid transit
buses [40, 41, 42]. e PEMS device collected mass emission
rates of CO2, CO, HC, O2, NOX, and particulate matter less
than 10μm (PM10) measurements at 1Hz. e CO2, CO, and
HC measurements are made using a nondispersive infrared
(NDIR) sensor, the O2 and NOX are measured with electro-
chemical sensors, and PM10 measurements made with a laser
scattering detector. GPS coordinates and engine parameters
from Controller Area Net work (CAN) signals available through
the vehicle’s OBD-II port can also be recorded. e emissions
were captured with an exhaust tailpipe probe, connected to t he
main dev ice with ruggedized tubing. A 12second delay correc-
tion was made in the PEMS to compile and synchronize all
emission data to account for delays from plumbing a nd sensors.
Lightning Hybrids Models
Lightning Hybrids in Loveland, CO retrots trucks, buses and
other large transit and delivery vehicles with hydraulic drive-
train-integrated systems. LH converts CVs to hydraulic hybrid
vehicles (HHV) that have complex energy management and
powertrain control systems. Customer drive cycles are not
always reproducible for FC and emissions research. is
provided an excellent opportunity to investigate the potential
of ANN hybrid vehicle FC and emissions modeling. A prelimi-
nary ANN approach was explored with real world, test track
data provided by Lightning Hybrids.
ANN Test and Training Procedures Due to t he preva-
lence of Matlab in control strategy research and the conve-
nience of the Neural Network Toolbox, Matlab was selected
as the modeling soware. A time series ANN was designed
and used to train all ANNs with a 3second time delay. e
time delay was included to account for short term operating
history of the vehicles such as the vehicle’s acceleration and
momentum. Preliminary tests showed a 3second delay yielded
more consistent ANNs over a 2second delay but showed no
discernable improvements for 4 to 5second delays. Vehicle
parameters and fuel consumption data was collected using
the LH system controller at 1Hz and emissions data collected
at 1Hz using the PEMS. Two distinct ANN models were
created for fuel consumption and emissions modeling. When
PEMS data were available, fuel consumption measured by the
PEMS was used instead of the LH controller measurement.
e two ANN models were trained with a varying set of input
variables as shown in Tab le 1 and Table 2 .
ANNs were trained using the Levenberg-Marquardt
backpropagation algorithm. Ten ANNs were trained for each
input scenario and the ANN with the most optimized R-value
and MSE was chosen. All drive cycle data was collected and
FIGURE 2 An example of a hydraulic hybrid retrofitted
truck from Lightning Hybrids.
© SAE International
TABLE 1 Fuel Consumption training input scenarios, Bare,
Vehicle Specific, and All In
FC Training Inputs Bare
Vehicle
Specific
(VS) All In
System State [Hybrid On/O] × × ×
Velocity [mi/hr] × × ×
Acceleration (time series derived)
[mi/hr/s]
× × ×
Engine Speed [rev/min] × ×
Arbitrary
Units
LH System High
Pressure
×
LH System Low Pressure ×
Throttle Position In ×
Brake Position ×
© SAE International
Downloaded from SAE International by Colorado State Univ, Wednesday, April 04, 2018

4 ECONOMIC AND EFFICIENT HYBRID VEHICLE FUEL ECONOMY AND EMISSIONS MODELING
© 2018 SAE International. All Rights Reserved.
provided by LH using vehicles with LH’s HHV system. Every
ANN model generated was tested and evaluated to determine
the absolute percent error for the FC and emissions for the
entire length of the tested drive cycle. Every model was tested
on data independent from training data.
To train the ANN model, four neuron amounts in the
hidden layer were used: 1, 3, 12, and 48. e Mass Para-Transit
Cycle was utilized for this exercise due to it having the longest
drive cycle. For all other ANNs trained, Eqn. 6 was used; this
method is supported by existing literature [50].
nNN
in out
=
-
2
(6)
where n is the number of neurons, Nin is the number of
unique inputs, and Nout is the number of unique outputs.
The fuel consumption ANNs were each tested for
performance among three training input scenarios as seen
in Table 1. e ANN models were trained on the Mass Para-
Transit Cycle again due to its length.
Previous work has not explored the inuence of the lengt h
of real world training data to create fuel consumption ANN
ts. e fuel consumption ANN models were trained with 5
datasets of varying length to show how training length aects
ANN performance.
Lightning Hybrids provided data for the Orange County
bus (OCBUS) and Mass Para-Transit Cycle fuel consumption
models and the European Delivery Cycle for emissions models.
e OCBUS cycle is dominated by lower cruising speeds (~20
mi/hr) and time spent at idle. e Mass Para-Transit Cycle has
periods of time spent at a stop with trends at intermediate to
high speeds (30 to 70 mi/hr). e European Delivery Cycle
depicts a delivery route and trends around intermediate speeds
(25 to 40 mi/hr). e OCBUS cycle is approximately 35minutes,
the Mass Para-Transit Cycle up to 9hours, and the European
Delivery Cycle is approximately 8minutes long.
Lightning Hybrids can change the state of their hybrid
vehicles (System on/o) during testing. is enabled a strong
comparison in models depicting either a conventional or a
hybrid vehicle. Drive data included system on and o control
setting and some included 2 dierent system-on settings:
Torque Addition (TA) and Torque Replacement (TR).TA adds
torque from the hybrid system to the conventional drivetrain
system. is reduces the driver’s need to increase the throttle
during operation. TR anticipates the throttle operation from
the driver and the hybrid system replaces some of the torque
in conventional drivetrain operation. is creates a more
conventional driving experience for the driver since the input
on the throttle should not dier between hybrid system on
and o.
e European Delivery Cycle was the only cycle tested
for the PEMS and was not explored extensively due to a limited
amount of data. e training input scenarios were explored
and are described in Table 2 .
Results
e number of neurons used to train the ANN did not seem
to aect the ANN t and produced small errors in the predic-
tions for fuel consumption. However, we did observe small
dierences in the errors as the number of neurons used to
train the ANN model were changed. ANNs trained with too
few or too many neurons performed poorly. e ANNs trained
with 1 and 48 neurons generated more error compared to the
3 neuron ANN which adhered to Eqn. 6 (Ta bl e 3). Figure 3
shows the 1 neuron ANN under predicting fuel consumption
at 30.5 min and 33min and the 48 neuron ANN under
predicting at 31.6min. While the 3 and 12 neuron ANNs
under or over predict in some instances, the 1 and 48 neuron
ANNs under or over predict more oen resulting in a larger
percent error over the length of the cycle.
For all tests, we used a ANN with one hidden layer that
included the number of neurons given by Eqn. 6. e number
of neurons seemed to have small eect on model performance
TABLE 2 Emissions training input scenarios, Bare, Vehicle
Specific, and All In
Emissions Training Inputs Bare
Vehicle
Specific All In
Velocity [mi/hr] × × ×
Acceleration (time series derived)
[mi/hr/s]
× × ×
Engine Speed (rev/min) × ×
ΔElevation (m) ×
© SAE International
TABLE 3 Percent error from testing Mass Para-Transit Cycle
trained ANNs and the approximate times to train each ANN.
4ANNs were trained with varying number of neurons in the
hidden layer and tested with the same Mass Para-Transit Cycle
Number of Neurons
in Hidden Layer
Fuel Conumption %
Error for 9 hour
Drive Cycle
Approximate Time
to Train (s)
1 0.23 % <5
3 0.12% 5
12 0.01 % 15
48 0.22 % 30
© SAE International
FIGURE 3 Comparisons of predicted fuel consumption
values from ANNs trained with varying amount of neurons in
the hidden layer.
© SAE International
Downloaded from SAE International by Colorado State Univ, Wednesday, April 04, 2018

ECONOMIC AND EFFICIENT HYBRID VEHICLE FUEL ECONOMY AND EMISSIONS MODELING 5
© 2018 SAE International. All Rights Reserved.
although our sensitivity studies suggest that the use of too few
(e.g. 1 or 3) and too many (e.g. 48) results in marginally higher
error. We hypothesize that too few or too many neurons may
create a weak ANN that cannot generalize (Figure 3). An
intermediate amount of neurons creates stronger performing
ANNs. e number of neurons determined by Eqn. 6 trains
an ANN that performs well but this equation may not work
for all applications. An ANN trained with 12 neurons
performs best among the 4 tested options but an equation or
method has not been found to support this for this study.
ANNs trained to predict fuel consumption all performed
well with absolute errors smaller than 1% for all training
scenarios. Models with more inputs beyond velocity, time,
and consequently acceleration, did not drastically improve
performance among these scenarios (Figure 4). Bare, VS, and
All In training scenarios yield 0.461%, 0.124%, and 0.974%
errors respectively. All subsequent fuel consumption ANNs
reported will be of the VS variety since it has the lowest error
in this case. Training ANNs with dierent input scenarios
shows that the Bare scenario that includes inputs of velocity
and acceleration are the most inf luential variables. The
addition of engine speed as an input variable only marginally
improves predictions of fuel consumption (Figure 4).
We examined the inuence of four dierent training
lengths from the OCBUS data on the ability of the ANN model
to predict fuel consumption. We  nd that the length of a dataset
used to train an ANN inuences the performance of the model.
Longer data sets produce sma ller errors and vice versa (Ta ble 4).
e length of a dataset that trains an ANN has signicant
eects on model performance. A substantial amount of data,
as well as increased variability in captured drive cycles,
improves the model. In Tabl e 4, ANNs with over 35minutes
of training data have error <1%. e ANN trained with
4minutes is the outlier with error of 1.50%, which produces
less error than the 8 and 17minute trained models.
NNs trained and tested with datasets from dierent
drive cycles perform depending on the conditions of the
training and testing datasets. ANN trained on a dataset
without highway driving: OCBUS cycle, and tested with data
with highway driving: Mass Para-Transit Cycle, yields a
13.08% error. An ANN trained on a dataset with highway
driving and tested with data without highway driving yields
a 4.01% error. Time of year and dierent trucks may inuence
these errors.
Variability in drive cycle also trains a more accurate
ANN. e OCBUS cycle, being an inner city bus route, does
not include higher speeds that would be found from driving
on the highway. e Mass Para-Transit Cycle trained ANN
model can predict the OCBUS cycle with error of 4.01% while
the OCBUS ANN model predicts the Mass Para-Transit Cycle
with a higher error of 13.08%.
“System on” trained ANNs can only accurately predict
“system on” fuel consumption and “system o” trained ANNs
can only predict “system o “fuel consumption. Large errors
(>10%) occur when a network trained on one system setting
attempts to predict fuel consumption from the opposite
system setting (Table 5). is indicates that an ANN model
does not model control system settings that it was not trained
to model.
e Lightning Hybrid system state can be set to more
than on and o settings. An ANN was built with the LH
system set to TA and tested with data collected with the system
set to TR. e TR prediction, 1.56% error, performs worse
than a TA prediction, 0.46% error, using a TA trained ANN.
e models performed well with training from dierent
LH system settings of controller on and o and torque addition
and torque replacement. In each case of ANNs built from data
with dierent system settings than the data used for testing,
predictions produced more error tha n when trained and tested
with the same settings (Tab le 5). ese instances of the models
failing shows that the trained ANNs are only trained for that
specic controller setting. Control system settings directly
affected performance of the ANN model predictions.
Extensive datasets that specify system settings would be able
to predict fuel consumption in each control scenario. is
would be invaluable in being able to determine the optimum
controller setting to reduce fuel consumption.
FIGURE 4 Predicted fuel consumption with varying
input scenarios compared to actual measured fuel
consumption values.
© SAE International
TABLE 4 ANNs trained with varying lengths OCBUS datasets
predict fuel consumption with associated percent errors.
Training Dataset Length (min) Fuel Consumption % Error
4 1.50%
8 3.30 %
17 2.24 %
35 0.46 %
70 0.12 %
© SAE International
TABLE 5 Percent error of aggregate fuel consumption from
training and testing ANNs with varying LH system on/o
settings. OCBUS cycle used for testing and training.
NN Trained Data Input Data % Error
LH System On LH System On 0.46 %
LH System On LH System O 14.81 %
LH System O LH System O 1.07%
LH System O LH System On 10.67%
© SAE International
Downloaded from SAE International by Colorado State Univ, Wednesday, April 04, 2018

6 ECONOMIC AND EFFICIENT HYBRID VEHICLE FUEL ECONOMY AND EMISSIONS MODELING
© 2018 SAE International. All Rights Reserved.
e ANNs trained to predict emissions perform well for
some outputs; notably well for HC and NOX with errors <5%
(Tab le 6). Predictions of CO consistently create error of >10%
for most scenarios.
NNs trained to predict emissions have higher error
and increasing the number of inputs does not appear to
improve performance (Tab le 6). This error may not be fully
attributed to the inputs. The European Delivery Cycle data
used to train ANNs is approximately 8minutes long; fuel
consumption ANNs trained on this length of data produced
more error in their predictions than any other tested data
length (Ta ble 4). A longer emissions data input could
benefit model performance. Separating each emission into
its own model may also improve model training since fuel
consumption models performed better overall while
predicting only one output.
Conclusions
In this study, the modeling of FC and emissions of custom
hydraulic hybrid vehicles was investigated using an ANN.
Parameters affecting model fidelity were investigated
including the number of neurons in the hidden layer,
specific training inputs, dataset length, and hybrid system
status. The results show FC errors within 0.1% and emis-
sions errors within 3% when proper ANN training
is utilized.
ANN are not only signicantly less computationally
costly than existing simulation standards such as Autonomie,
but they are potentially more accurate. is demonstrates
signicant benet for designing improved vehicle control
strategies (Eco-Driving, Optimal Energy Management),
reducing the need for physical vehicle testing, accurately
capturing the emissions results from slight drive cycle varia-
tions, improving the understanding of real world emissions
and fuel impacts, and enabling high delity learning control
in physical vehicles.
Our work does not consider changes in FC and emissions
from variations in fuel, engine operation over time, driver
behavior, and ambient conditions, to name a few. Although
important, consideration of these variables were beyond the
scope of the outlined work, which was to demonstrate the
applicability of a neural network model to predict fuel
consumption and emissions. This need to explored in
future work.
References
1. Energy Information Administration, “Energy Use for
Transportation,” http://www.eia.gov/
Energyexplained/?page=us_energy_transportation#tab1,
accessed 5 Oct 2016.
2. International Energy Agency. “Key World Energy Statistics
2016,” International Energy Agency, 2015.
3. U.S. EPA, “Integrated Science Assessment for Oxides of
Nitrogen - Health Criteria (Second External Review Dra,
2015),” EPA/600/R-14/006, (Washington, DC, U.S.
Environmental Protection Agency, 2015).
4. U.S. EPA, “Final Assessment: Integrated Science Assessment
for Carbon Monoxide”, EPA/600/R-09/019F. (Washington,
DC, U.S. Environmental Protection Agency, 2010).
5. Kanakidou, M., Seinfeld, J.H., Pandis, S.N., Barnes, I. et al.,
“Organic Aerosol and Global Climate Modelling: A Review,”
Atmospheric Chemistry and Physics 5(4):1053-1123, 2005.
6. Kim, K.-H., “A Review of Airborne Polycyclic Aromatic
Hydrocarbons (PAHs) and their Human Health Eects,”
Environment International 60:71-80, 2013.
7. Harrison, R.M. and Yin, J., “Particulate Matter in the
Atmosphere: Which Particle Properties Are Important for its
Eects on Health?” Science of the Total Environment
249(1):85-101, 2000.
8. Jiao, W., Christopher Frey, H., and Ye, C., “Assessment of
Inter-Individual, Geographic, and Seasonal Variability in
Estimated Human Exposure to Fine Particles,”
Environmental Science & Technology 46(22):12519-
12526, 2012.
9. U.S. EPA, “Final Report: Integrated Science Assessment for
Particulate Matter”, EPA/600/R- 08/139F. (Washington, DC,
U.S. Environmental Protection Agency, 2009).
10. U.S. Department of Transportation, “Corporate Average Fuel
Economy (CAFE) Standards,” https://www.transportation.
gov/mission/sustainability/corporate-average-fuel-economy-
cafe-standards, accessed January 17, (2017).
11. Asher, Z., Wifvat, V., Navarro, A., Samuelsen, S. et al., “e
Importance of HEV Fuel Economy and Two Research Gaps
Preventing Real World Implementation of Optimal Energy
Management,” SAE Technical Paper 2017 26 010 6, 2017,
doi:10.4 271/2017-26 -0106.
12. Liu, H. and Barth, M., “Identifying the Eect of Vehicle
Operating History on Vehicle Running Emissions,”
Atmospheric Environment 59:22-29, 2012.
TABLE 6 Percent error of each predicted emission type over the length of the European Delivery Cycle for system on and o and
the 3 input scenarios.
Contra System
Setting System On System O
Emission % Error Bare VS All In Bare VS All In
CO20.87 % 2.13% 7.2 6 % 1.29% 4 11 % 11.31 %
CO 6.92 % 62.63 % 15.33% 13.61 % 2.73 % 41.50 %
HC 1.72 % 2.02 % 0.27 % 0.27 % 1.91 % 1.38%
NOx0.15 % 0.89 % 2.51 % 2.95 % 2.14 % 3.15%
PM10 0.37 % 0.48 % 7.23 1.59% 2.00 % 6.06 %
© SAE International
Downloaded from SAE International by Colorado State Univ, Wednesday, April 04, 2018

ECONOMIC AND EFFICIENT HYBRID VEHICLE FUEL ECONOMY AND EMISSIONS MODELING 7
© 2018 SAE International. All Rights Reserved.
13. Ahn, K., Rakha, H., Trani, A., and Van Aerde, M.,
“Estimating Vehicle Fuel Consumption and Emissions
Based on Instantaneous Speed and Acceleration Levels,”
Journal of Transportation Engineering 128(2):182-
190, 2002.
14. U.S. EPA, “Development of Emission Rate for Light Duty
Vehicle in the Motor Vehicle Emissions Simulator
(MOV ES 2010), EPA-420-R-11-011,” (Ann Arbor, MI, U.S.
Environmental Protection Agency, 2010).
15. Abou-Senna, H. and Radwan, E., “VISSIM/MOVES
Integration to Investigate the Eect of Major Key Parameters
on CO 2 Emissions,” Transportation Research Part D:
Transport and Environment 21:39-46, 2013.
16. Wallace, H.W., Jobson, B.T., Erickson, M.H., McCoskey, J.K.
et al., “Comparison of Wintertime CO to NOx Ratios to
MOVES and MOBILE6.2 on-Road Emissions Inventories,”
Atmospheric Environment 6 3:2 89-297, 2 01 2 .
17. Zhang, Y., LV, J., and Wang, W., “Evaluation of Vehicle
Acceleration Models for Emission Estimation at an
Intersection,” Transportation Research Part D: Transport and
Environment 18:4 6-50, 2013.
18. Choi, D. and Koupal, J., “MOVES Validation,” MOVES
Workshop, Ann Arbor, 14-16, June 2011.
19. Liu, B. and Christopher Frey, H., “Variability in Light-Duty
Gasoline Vehicle Emission Factors from Trip-Based Real-
World Measurements,” Environmental Science & Technology
49(20):12525-12534, 2015.
20. Pelkmans, L. and Debal, P., “Comparison of On-Road
Emissions with Emissions Measured on Chassis
Dynamometer Test Cycles,” Transportation Research Part D:
Transport and Environment 11(4):233-241, 2006.
21. Eciency, Energy, and Renewable Energy, “Evaluation of the
2010 Toyota Prius Hybrid Synergy Drive System,”
https://pdfs.semanticscholar.org/a5d4/3e8a28ae3bc3297ed44
9f18b42c817083 9.pdf, 2011.
22. Kim, N., Duoba, M., Kim, N., and Rousseau, A., “Validating
Volt PHEV Model with Dynamometer Test Data Using
Autonomie,” SAE Int. J. Passeng. Cars - Mech. Syst. 6(2):985-
992, 2013, doi:10.4 271/2013-01-1458.
23. Rajamani, R., “Vehicle Dynamics and Control,” (Springer
Science & Business Media, 2011).
24. Asher, Z.D., Baker, D.A., and Bradley, T.H., “Prediction
Error Applied to Hybrid Electric Vehicle Optimal Fuel
Ec onomy,” IEEE Transactions on Control Systems Technology
PP 99: 1-14, 2017.
25. Layne, J.R., Passino, K.M., and Yurkovich, S., “Fuzzy
Learning Control for Antiskid Braking Systems,” IEEE
Transactions on Control Systems Technology 1(2):122-
129, 19 93.
26. Nelles, O., “IntelligenTip®: A Learning Driving Strategy for
Automated Transmissions,” SAE Technical Paper 2003- 01-
0534, 2003, doi:10.4271/2003-01-0534.
27. Asher, Z.D., Trinko, D.A., and Bradley, T.H., “Increasing the
Fuel Economy of Connected and Autonomous Lithium Ion
Electried Vehicles, Chapter 6, Behaviour of Lithium-Ion
Batteries in Electric Vehicles: Battery Health, Performance,
Safety, and Cost,” 1st Ed, (Springer International Publishing,
2018), in press.
28. Demuth, H.B., Beale, M.H., De Jess, O., Hagan, M.T. “Neural
Network Design”. (USA, Martin Hagan, 2014)
29. Rosenblatt, F., “e Perceptron: A Probabilistic Model for
Information Storage and Organization in the Brain,” Psychol.
Rev. 65:386-408, 1958.
30. Hopeld, J.J., “Neural Networks and Physical Systems with
Emergent Collective Computational Abilities,” Proc. Natl.
Acad. Sci. U. S. A. 79:2554-2558, 1982.
31. Rumelhart, D.E., McClelland, J.L., and PDP Research Group,
C, editors, “Parallel Distributed Processing: Explorations in
the Microstructure of Cognition, Vol. 1,” (Cambridge,
Foundations MIT Press, 1986).
32. Sou, M.D., Ghobadian, B., Naja, G., Sabzimaleki, M., and
Jaliliantabar, F., “Performance and Exhaust Emissions of a SI
Two-Stroke Engine with Biolubricants Using Articial
Neural Network,” Energy Procedia 75:3-9, 2015, doi:10.1016/j.
egypro.2015.07.127.
33. Çay, Y., Korkmaz, I., Çiçek, A., and Kara, F., “Prediction of
Engine Performance and Exhaust Emissions for Gasoline
and Methanol Using Articial Neural Network,” Energy
50:177-186 , 2013, doi:10.1016/j.energy.2012.10.052.
34. Ashhab, M.S.S., “Fuel Economy and Torque Tracking in
Camless Engines through Optimization of Neural
Networks,” Energy Convers Manag 49:365-372, 2008,
doi:10.1016/j.enconman.2007.06 .005.
35. Sayin, C., Ertunc, H.M., Hosoz, M., Kilicaslan, I. et al.,
“Performance and Exhaust Emissions of a Gasoline Engine
Using Articial Neural Network,” Appl erm Eng 27:4 6-54,
2007, doi :10.1016/j.applther ma leng. 20 06.05.016.
36. Li, H., Butts, K., Zaseck, K., Liao-McPherson, D. et al.,
“Emissions Modeling of a Light-Duty Diesel Engine for
Model-Based Control Design Using Multi-Layer Perceptron
Neural Networks,” SAE Technical Paper 2017 01 0601, 2017,
doi:10.4 271/2017-01-0601.
37. Turkson, R.F., Yan, F., Ali, M.K.A., and Hu, J., “Articial
Neural Network Applications in the Calibration of Spark-
Ignition Engines: An Overview,” Eng Sci Technol Int J
19:1346 -1359, 2016, doi:10.1016/j.jestch.2016.03.003.
38. Nielsen, M.A., “Neural Networks and Deep Learning,”
http://neuralnetworksanddeeplearning.com, 2015.
39. Documentation. Retrieved 5 Oct 2016, URL:http://www.
mathworks.com/help/nnet/gs/neural-network-time-series-
prediction-and-modeling.html, (n.d.).
40. Liu, B. and Christopher Frey, H., “Variability in Light-Duty
Gasoline Vehicle Emission Factors from Trip-Based Real-
World Measurements,” Environmental Science & Technology
49(20):12525-12534, 2015.
41. Frey, H.C., Zhang, K., and Rouphail, N.M., “Fuel Use and
Emissions Comparisons for Alternative Routes, Time of
Day, Road Grade, and Vehicles Based on in-Use
Measurements,” Environmental Science & Technology
42(7):2483-2489, 2008.
42. Sandhu, G.S., Christopher Frey, H., Bartelt-Hunt, S., and
Jones, E., “In-Use Measurement of the Activity, Fuel Use,
and Emissions of Front-Loader Refuse Trucks,” Atmospheric
Environment 92:557-565, 2014.
43. Hallmark, S.L., Bo, W., and Sperry, R., “Comparison of on-
Road Emissions for Hybrid and Regular Transit Buses,”
Downloaded from SAE International by Colorado State Univ, Wednesday, April 04, 2018

8 ECONOMIC AND EFFICIENT HYBRID VEHICLE FUEL ECONOMY AND EMISSIONS MODELING
Journal of the Air & Waste Management Association
63(10):1212-1220, 2013.
44. Frey, H.C., Zhang, K., and Rouphail, N.M., “Vehicle-Specic
Emissions Modeling Based upon on-Road Measurements,”
Environmental Science & Technology 44(9):3594-3600, 2010.
45. Lee, T. and Christopher Frey, H., “Evaluation of
Representativeness of Site-Specic Fuel-Based Vehicle
Emission Factors for Route Average Emissions,”
Environmental Science & Technology 46(12):6867-6873, 2012.
46. Frey, H.C., Yazdani-Boroujeni, B., Hu, J., Liu, B. et al. “Field
Measurements of 1996 to 2013 Model Year Light Duty
Gasoline Vehicles,” Proceedings, 106th Annual Conference,
Air & Waste Management Association, Chicago, IL, 2013.
47. Barlow, T.J., Latham, S., McCrae, I.S., and Boulter, P.G. “A
Reference Book of Driving Cycles for Use in the
Measurement of Road Vehicle Emissions.” TRL Published
Project Report, 2009.
48. Global MRV, “PEMS AXIONR/S+: PDF Brochure,” htt p: //
www.globalmrv.com/pems-axionrs-2/, accessed 17 Jan 2017.
49. Franco, V., Kousoulidou, M., Muntean, M., Ntziachristos, L.
et al., “Road Vehicle Emission Factors Development:
AReview,” Atmos Environ 70 :84- 97, 2013 .
50. Heaton, J., “Introduction to Neural Networks with Java,”
(Heaton Research, Inc., 2008).
51. Galang, A., “Predicting Hybrid Vehicle Fuel Economy and
Emissions with Neural Network Models Trained with Real
World Data,” Master of Science, Colorado State University,
Libraries, https://dspace.library.colostate.edu/
handle/10217/181369, 2 017.
Contact Information
omas H. Bradley, Ph.D.
Colorado State University, Fort Collins, CO 80523 USA
omas.Bradley@ColoState.edu
Definitions/Abbreviations
ANN - Articial Neural Network
CAFE - Corporate Average Fuel Economy
CV - Conventional Vehicle
EPA - Environmental Protection Agency
FE - Fuel Economy
GHG - Greenhouse Gas Emissions
HC - hydrocarbons
HEV - Hybrid Electric Vehicle
MOVES - MOtor Vehicle Emission Simulator
NARX - Nonlinear AutoRegressive network with eXogenous
PEMS - Portable Emissions Monitoring System
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, o r transmitted, in any form or by any means , electronic, mechanical,
photocopying, recording, or otherwise, without the prior written permission of th e copyright holder.
Positions and opinions advanced in this paper are those of the author(s) and not ne cessarily those of SAE International. The author is solely responsible for the
content of the paper.
ISSN 0148-7191
Downloaded from SAE International by Colorado State Univ, Wednesday, April 04, 2018