Energy and AI 6 (2021) 100115

Contents lists available at ScienceDirect

Energy and AI
journal homepage: www.sciencedirect.com/journal/energy-and-ai

Constructing representative driving cycle for heavy duty vehicle based on

Markov chain method considering road slope
Xinyi Jia, Hewu Wang, Liangfei Xu *, Qing Wang, Hang Li, Zunyan Hu, Jianqiu Li *,
Minggao Ouyang
State Key Lab of Automotive Safety and Energy, School of Vehicle and Mobility, Tsinghua University, Beijing 100084, China

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• An effective road slope observation al­

gorithm is designed based on GPS
signals.
• A 2D Markov Chain method together
with an average speed-based matching
algorithm is proposed for constructing
representative driving cycles with road
slope.
• The necessity of considering road slope
in constructing driving cycles is vali­
dated based on real data.

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

Key words: Electrification of heavy duty vehicles (HDVs) is critical to realization of the target of carbon neutralization in the
Driving cycles future. For most HDVs, the influence of road slope on vehicle power usually cannot be ignored due to significant
Markov chain method road slope variation during long driving mileages. In order to design the powertrain system for electrified HDVs
Power demand
effectively, it is necessary to construct representative driving cycles with road slope information. There are two
Road slope observation
Segment matching
difficulties for this task. (1) Road slope measuring devices are usually costly. A cheaper yet effective method for
measuring road slope needs to be developed. (2) A 3D (three dimension) Markov chain method is usually utilized
for constructing cycles with velocity and road slope. This method is complex and time consuming, and needs to
be improved. In this paper, a 2D (two dimension) Markov chain method for addressing these issues is proposed. A
road slope observation is designed based on normal GPS (Global Positioning System) signals and a high order
Butterworth filter. The effectiveness of the method is validated. Driving velocity and road slope are collected and
observed for the area between Beijing and Zhangjiakou in northern China. Representative cycles with road slope
are constructed using a 2D Markov chain method and a matching algorithm based on average speed. With the
introduced technology, three representative driving cycles with road slope for urban, suburban and highway
routes are designed. Statistic results on vehicle power show that, the representative driving cycles are effective
with relative errors less than 4% compared to the real driving conditions. These driving cycles will be utilized in
designing electric HDVs, such as hydrogen fuel cell vehicles in the future.

* Corresponding authors: State Key Lab of Automotive Safety and Energy, School of Vehicle and Mobility, Tsinghua University, Beijing, 100084, China.
E-mail addresses: xuliangfei@tsinghua.edu.cn (L. Xu), lijianqiu@tsinghua.edu.cn (J. Li).

https://doi.org/10.1016/j.egyai.2021.100115
Received 18 July 2021; Received in revised form 12 September 2021; Accepted 13 September 2021
Available online 26 September 2021
2666-5468/© 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
X. Jia et al. Energy and AI 6 (2021) 100115

1. Introduction

At present, with the increasingly serious energy and environmental

problems, heavy duty vehicles (HDVs) have attracted great attentions
for their potential to reduce fuel consumption and green-house-gas
(GHG) emission [1-2]. Among different technologies for vehicle elec­
trification, proton exchange membrane fuel cell (PEMFC) vehicles have
become the focus of the automotive industry for their high efficiency and
zero emissions [3-5]. However, the durability of fuel cells remains key
constraint to the large-scale commercialization [6-8]. In order to design
suitable fuel cell powertrains for HDVs, it is necessary to investigate
representative driving cycles in typical operating areas [9]. Fig 2. Test Device. (a) Columbus V990 GPS Logger. (b) gyroscope RT3000.
A representative driving cycle has similar statistic characteristics
with the actual driving condition [10-12]. Such driving cycles are the
basis for vehicle design [13-14]. Driving cycles are developed to eval­ Table 1
uate fuel consumption and emission performance of vehicles under Characteristic parameters definition.
real-world driving conditions [15-16]. Based on this, energy manage­
Parameters Definitions
ment strategy and power system configuration can be optimized [17].
Speed Parameters Average Speed Vm Average Speed including idle speed
Driving conditions are complex and can be affected by geography, traffic
Average Driving Average Speed excluding idle speed
conditions, drivers’ behaviors and other factors [18-20]. Different
Speed Vmr
driving cycles are constructed for different regions, different vehicle
Maximum Speed Maximum Speed
types, and different application scenarios [21-22]. Vmax
Researchers have proposed many methods to construct driving cy­ Acceleration Average Average Acceleration
cles from actual driving data, to make driving cycles close to actual Parameters Acceleration am
condition. Markov Chain method is a common and effective method for Average Average Deceleration
driving cycles construction [23-24]. The Markov Chain method was Deceleration dm
firstly used in vehicle driving cycles by Kolmanovsky et al. for the Time Ratio Idle Ratio pi Proportion of idle time( − 0.15 < a <
optimization of automotive powertrain strategies [25]. The application 0.15m/s2 , v < 0.5km/h)
of the Markov Chain method in driving cycles has been gradually Acceleration Ratio Proportion of acceleration
improved and has become quite widely used afterwards. Yang et al. pa time(a ≥ 0.15m/s2 )
applied this method to construct a representative driving cycle for the Deceleration Proportion of deceleration time(a ≤ −
Ratio pd 0.15m/s2 )
EV (Electric Vehicles) in Shenyang, China [26]. Wang et al. built the
driving cycles of the taxies in Beijing [27]. Markov method is further Cruise Ratio pr Proportion of cruise time( − 0.15 < a <
0.15m/s2 , v ≥ 0.5km/h)
combined with PCA, K-means and support vector machine [28]. In these
studies, only two states (vehicle speed and acceleration) were utilized in
constructing representative driving cycles, so it is normally called as the
[30-31]. Gwenaëlle Souffran et al. used three states to construct driving
2D Markov Chain method.
cycles, e.g., the speed, acceleration, and road slope [32], which is nor­
For HDVs, the impact of road slope is not ignorable, because such
mally named as the 3D Markov method. This method is effective in
vehicles normally have long driving mileages and the road slope changes
constructing driving cycles with road slope, but it is complex and re­
significantly. The road slope affects the driving force and power de­
quires long computing time. A simple yet effective method for con­
mand, and needs to be considered in powertrain design [29]. In litera­
structing driving cycles with road slope is strongly required.
ture, most studies on driving cycles focus on the 2D Markov chain
In this paper, a new method is proposed to construct typical driving
method, and few studies take the effects of road slope into account

Fig 1. Process of constructing driving cycles with road slope.

2
X. Jia et al. Energy and AI 6 (2021) 100115

Table 2
Characteristic parameters of three different routes.
Parameters Urban Suburban Highway
Route Route Route

Average Speed Vm (km /h) 19.92 22.64 51.16

Average Driving Speed 27.32 31.61 65.47
Vmr (km /h)
Maximum Speed Vmax (km /h) 57 85.71 99.50
Average Acceleration 0.4947 0.3637 0.3591
am (m /s2 )
Average Deceleration -0.5488 -0.5308 -0.3968
dm (m /s2 )
Idle Ratio pi (%) 27.21 28.40 17.55
Acceleration Ratio pa (%) 27.00 28.66 24.26
Deceleration Ratio pd (%) 23.52 19.57 21.82
Cruise Ratio pr (%) 22.31 23.34 36.38

Fig 3. State transition matrix.

cycles for HDVs such as fuel cell vehicles. Three typical driving cycles
with road slope for urban, suburban and highway routes are designed
using this technology. The paper is organized as follows: Section 2 in­
troduces the new method in general. Section 3 presents the construction
of the driving cycles without slope. Section 4 introduces how to integrate
the slope into the driving cycles. Section 5 is the conclusion.

2. Introduction to the new method for constructing driving

cycles with road slope

In order to get an optimized powertrain design of HDVs, it is

necessary to construct a representative driving cycle with road slope.
The vehicle velocity, acceleration and road slope are treated as three
state variables, and normally a 3D Markov Chain method is needed to Fig 4. Representative driving cycles without road slope. (a)urban route. (b)
suburban route. (c)highway route.
solve the problem. Since the 3D Markov Chain method is complex and
time consuming, a new method is proposed . The new method for con­
structing driving cycles with road slope is based on a 2D Markov chain divided into several segments, which are all labeled using corresponding
method, road slope observation and a matching algorithm. The process average speed of the driving cycle. Meanwhile, the observed road slope is
of the new method can be conducted in three steps shown in Fig. 1. also divided into several segments with same interval and is labeled with
In Step 1, a typical driving cycle without road slope is constructed corresponding average speed of the real driving conditions. Considering the
using a normal 2D Markov chain method. Meanwhile, the speed and coupling effect between vehicle velocity and road slope, a segment match­
altitude signals are collected by a GPS (Global Positioning System) ing algorithm is proposed to construct a suitable road slope sequence cor­
equipment, and are utilized to observe the road slope instead of using responding to a representative driving cycle. For each driving cycle
extra expensive devices. segment, a road slope segment which is labeled with a nearest average speed
In Step 2, a driving cycle with road slope is constructed according to a is selected. By splicing the road slope segment together, a new driving cycle
matching algorithm based on average speed. The driving cycle in Step 1 is with road slope is then constructed.

3
X. Jia et al. Energy and AI 6 (2021) 100115

Table 3
Comparison of characteristic parameters.
Parameters Urban Route Suburban Route Highway Route
Real data Driving cycles Relative errors Real data Driving cycles Relative errors Real data Driving cycles Relative errors

Vm (km /h) 19.92 20.02 0.50% 22.64 21.67 -4.28% 51.16 49.91 -2.44%
Vmr (km /h) 27.32 26.49 -3.04% 31.61 30.10 -4.78% 65.47 60.89 -7.00%
Vmax (km /h) 57.00 58.00 1.75% 85.71 83.00 -3.16% 99.50 97.40 -2.11%
am (m /s2 ) 0.4947 0.4993 0.93% 0.3637 0.3519 -3.24% 0.3591 0.3507 -2.34%
dm (m /s2 ) -0.5488 -0.5104 -7.00% -0.5308 -0.5332 0.45% -0.3968 -0.4110 3.58%
pi (%) 27.21 24.43 -10.22% 28.40 28.00 -1.41% 17.55 17.98 2.45%
pa (%) 27.00 25.07 -7.15% 28.66 30.06 4.88% 24.26 22.33 -7.96%
pd (%) 23.52 24.86 5.70% 19.57 20.44 4.45% 21.82 19.22 -11.92%
pr (%) 22.31 25.64 14.93% 23.34 21.50 -7.88% 36.38 40.47 11.24%

3. Constructing driving cycles without road slope

3.1. Route Selection and data collection

In order to reveal the driving characteristics of HDVs in typical areas,

three representative routes around Zhangjiakou are selected, including
an urban, a suburban and a highway route. Buses are equipped with the
Columbus V990 GPS Logger, as shown in Fig. 2(a), to record date, time,
latitude, longitude, altitude and speed at a sample frequency of 1Hz. The
resolutions of altitude and speed are 1m and 1km/h respectively. In
order to verify the effectiveness of the road slope observation algorithm,
a road slope sensor equipment, the gyroscope RT3000, is equipped as
shown in Fig. 2(b).

3.2. Data preprocessing and key parameters extraction

Due to external interference or internal problems, signal loss and

errors may occur during data recording process. The abnormal data,
such as extremely high or low values of speed or acceleration, are
Fig 5. Relative errors of key characteristic parameters. removed and replaced during the data preprocessing.
The vehicle driving condition can be described with characteristic
In Step 3, the effectiveness of driving cycles with road slope is veri­ parameters. In this study, nine characteristic parameters are selected for
fied by comparing the statistic characteristics of vehicle power between driving cycle construction. These characteristic parameters are mainly
the real driving road conditions and the constructed driving cycles. divided into three categories, which reflect characteristics of speed,
acceleration, and time ratio of vehicle status, respectively. The speed
parameters include average speed, average driving speed and maximum
speed. The acceleration parameters include average acceleration and

Fig 6. Altitude changes on different routes. (a)urban route. (b)suburban route. (c)highway route.

4
X. Jia et al. Energy and AI 6 (2021) 100115

3.3. Driving cycles based on 2D Markov chain method without road slope

As mentioned in Fig. 1, representative driving cycles without road

slope are firstly developed using a 2D Markov Chain method. Markov
Chain is a stochastic process with Markov property and has either
discrete state space or a discrete index set [34]. Markov property is that,
the probability of one current state Xt changing to the next state Xt+1 is
only related to the current state and independent from the previous
states Xt-1,… , X1, i.e. the conditional probability p of the state variable X
satisfies Eq. (1) [35].
p(Xt+1 Xt , …, X1 ) = p(Xt+1 Xt ) (1)
In Markov chains, the probability from the current state Xi to the next
state Xj is defined as the transition probability pij , which can be calcu­
lated statistically. The transition probability of each state can be
calculated, and all the transition probabilities constitute the state tran­
sition matrix (TPM) as:
⎡ ⎤
[ ] p11 ⋯ p1I
TPM = pij ij = ⎣ ⋮ ⋱ ⋮ ⎦ (2)
Fig 7. Comparison of calculated slope and measured slope. pI1 ⋯ pII I×I

where I is the number of the states; pij is obtained from the collected
data.
The driving conditions (speed and acceleration) can be considered as
a stochastic process [36]. The speed and acceleration for next moment
are solely dependent on the speed and acceleration for current moment,
and conditionally independent on all other moments. Based on this, the
driving cycles with vehicle speed can be developed using the Markov
Chain method. The modeling method is performed in three steps:
Step1: Constructing the Markov state. In the 2D Markov Chain
method, the vehicle velocity and acceleration are regarded as two in­
dependent state variables. The vehicle speed range is discretized into
several speed segments vi with an interval of 1km/h. The acceleration
range is also discretized into several segments aj with an interval of
0.1m/s2 .
Step2: Calculating the state transition probability matrix. With the
real on-road data, the state transition matrix is constructed based on the
transition probabilities in Eq. (3). According to this equation, when the
current speed belongs to the segment vi and the current acceleration
belongs to the segment aj , the probability of the next speed belongs to
the segment vk can be calculated as follows.
( ) N(i, j, k)
Pijk = P vt+1 ϵvk vt ϵvi , at ϵaj = (3)
N(i, j)
N(i, j, k) is the frequency when the current speed belongs to vi , the
current acceleration belongs to aj , and the next speed belongs to vk .
N(i, j) is the total frequency of occurrences when the current speed be­
longs to vi and the current acceleration belongs to aj . In this way, by
counting all the possible situations, a complete state transition matrix is
Fig 8. Observed road slope of three real driving routes. (a)urban route. (b) established as in Fig. 3.
suburban route. (c)highway route. Step3: Constructing driving cycles. With the initial value of speed
and acceleration v0 = 0, a0 = 0, the next time step of speed v1 is
average deceleration. The time ratio parameters mainly reflect the randomly generated according to the distribution of state transition
proportion of different driving modes, including idle ratio, acceleration matrix. The corresponding acceleration a1 is calculated by v0 and v1 .
ratio, deceleration ratio and cruise ratio. Definition of each character­ The next step of speed v2 and acceleration a2 are calculated in the same
istic parameter refer to the national standard is listed in Table 1[33]. way. The above steps are repeated continuously for the whole time
The characteristic parameters of three routes are extracted from real range, and the representative driving cycle with vehicle velocity infor­
data and the results are shown in Table 2. The characteristic parameters mation is constructed. The time range of the driving cycle is set to be
have significant distinctions among three road conditions in terms of 1400s for the urban route, 1800s for the suburban and highway routes
speed, acceleration and time ratio. The driving condition of urban route according to the national standard. The generated representative driving
shows low speed, high acceleration/deceleration and high time ratio of cycles of urban route, suburban route and highway route are shown as
acceleration/deceleration. The highway route is characterized by high Fig. 4 (a) (b) (c), representatively.
speed, low acceleration/deceleration and high proportion of cruise
driving mode. The parameter values of suburban route are moderate.

5
X. Jia et al. Energy and AI 6 (2021) 100115

Fig 9. Segment matching algorithm.

3.4. Relative errors of characteristic parameters function of the Butterworth filter is stated in Eq. (6).
1
Characteristic parameters are the criteria for judging whether a |H(jω)|2 = ( )2n (6)
driving cycle can represent the actual driving situation in terms of 1+ ω
ωc
vehicle velocity. The characteristic parameters extracted from the actual
driving data are shown in Table 2, and the comparison with those of The key parameters are cut-off frequency ωc and order n. The pass
representative driving cycles are listed in Table 3. Relative errors of key band and stop band cut-off frequency is designed referring to the power
parameters are shown in Fig. 5. spectral density. The filter parameters are optimized using real data.
Results show that, the characteristic parameters of generated driving Finally, a 14th order Butterworth filter with normalized pass and stop
cycles are in good agreement with those of actual data. In particular, the band cut-off frequencies of 0.1 and 0.3 is utilized.
relative errors of the speed and acceleration characteristic parameters To validate the effectiveness of this method, the gyroscope RT3000,
are basically less than 7%. The errors of the time ratio parameters are as mentioned before, is utilized to measure the road slope in a specific
relatively large, but are still less than 15%. Since speed and acceleration route around Zhangjiakou city. The comparison of the calculated and
information are more important than time ratio parameters for the the measured road slope is shown in Fig. 7, indicating that the estimated
driving cycles, such relative errors are acceptable. These results indicate road slope can approximate the actual road slope.
that, the constructed driving cycles well reflect the real operating con­ The curves of road slope of the three real routes, i.e. urban, suburban
ditions in terms of vehicle velocity. and highway routes around Zhangjiakou city as in Fig. 6, are shown in
Fig. 8.
4. Constructing driving cycles with road slope

4.1. Road slope observation and validation 4.2. Segment matching algorithm for velocity and road slope segment

Road slope has a great influence on the power demand of HDVs. In For the three typical driving routes around Zhangjiakou city, i.e.
the demonstration area around Zhangjiakou, the altitude changes urban, suburban and highway routes as shown in Fig. 6, the represen­
significantly, especially on intercity routes. The curves indicating alti­ tative driving cycles (Fig. 4) and observed road slope for real driving
tude changes on urban, suburban and highway routes are shown in conditions (Fig. 8) are obtained. The driving cycles with vehicle speed in
Fig. 6. Fig. 4 and road slope in Fig. 8 have different time bases, and cannot be
In order to obtain the information of road slope in the demonstration combined directly. Vehicle speed and road slope are coupled with each
area, a road slope observation based on GPS signals is designed. The other, i.e. road slope affects drivers’ behaviors and vehicle speed.
road slope is calculated as shown in Eqs. (4) and (5): integrating velocity Considering the coupling effect between vehicle velocity and road slope,
for the driving distance s(tm ) and getting the slope angle α from the a segment matching algorithm is proposed to construct a suitable road
change of altitude and driving distance. slope sequence corresponding to a representative driving cycle, without
increasing computation time.
s(tm ) = v(t)Δt (4) The detailed segment matching algorithm is shown in the Fig. 9. It is
observed that the road slope does not change much within 10 seconds.
1 Δh Therefore, both the representative driving cycles in Fig. 4 and road slope
α = sin− (5)
Δs sequences in Fig. 8 are divided into several segments with a time interval
Δt is the sample time period, v(t) is the sampled vehicle velocity, Δh of 10 seconds. Each speed segment is labeled by its average speed vv (i),
is the change of altitude during the sample time period, Δs is the change and each road slope segment S(i) is also labeled by the corresponding
of driving distance during sample time period. The original value of road actual average speed vs (i). For each driving cycle speed segment, a road
slope angle α contains lots of noises. In order to remove high-frequency slope segment which is labeled with a nearest average speed is selected.
noises, a Butterworth low-pass filter is utilized [37]. The transfer That is to say, assuming the driving cycle speed segment V(i) is labeled
with an average speed vv (i), the matching algorithm will search all the

6
X. Jia et al. Energy and AI 6 (2021) 100115

Table 5
Power Demand Comparison between Driving Cycles and Actual Routes.
Urban Suburban Highway
Route Route Route

Average driving power with observed 64.9712 56.6442 107.1398

road slope on actual driving routes
(kW)
Average driving power with observed 62.4884 56.0747 104.0215
road slope on representative driving
cycles (kW)
Average driving power ignoring road 52.2706 49.8077 83.4950
slope on actual driving routes (kW)
Relative error of driving power causing 24.30% 13.73% 28.32%
ignoring road slope
Relative error of driving power causing -3.82% -1.01% -2.91%
using the 2D Markov chain method
with road slope

road slope segments and find the optimal one S(k) with a minimal
average speed error, i.e. min|vv (i) − vs (k)| (k∈{1,2,…, N}). Then, a new
driving cycle segment D(i), which is composed of a corresponding
driving cycle segment V(i) and slope segment S(k), is created. Traversing
i from 1 to the end, a driving cycle with road slope information can be
spliced.

4.3. Driving cycles with road slope based on segment matching

The three driving cycles with road slope are constructed using the
segment matching algorithm referred above, as shown in Fig. 10.

4.4. Relative errors of power requirement

To verify the impacts of road slope on driving power and whether the
driving cycles with slope can accurately reflect the actual vehicle power
demand, the driving power of typical fuel cell buses are calculated. Basic
Eqs. for driving power are shown in Eqs. (7) and (8).
1
F = δma + CD Aρv2 + mgf cosα + mgsinα (7)
2

Pbus = F⋅v (8)

Pbus is the driving power, F is the driving force of the bus, including
the inertial force δma, wind resistance 12CD Aρv2 , friction resistance
mgfcosα and climbing resistance mgsinα. Symbols a, α and v represent
respectively the acceleration, slope angle and speed. The definitions and
values of other bus parameters are listed in Table 4.
The average driving power demands on actual routes (Fig. 6 and
Fig. 8) and representative driving cycles (Fig. 10) are compared in
Table 5. For the real three typical driving routes, the average driving
power with road slope are 64.97kW (urban), 56.64kW (suburban) and
107.14kW (highway), respectively. Using the 2D Markov chain method
Fig 10. Representative driving cycles and corresponding slope. (a)urban route. and matching algorithm based on average speed, three representative
(b)suburban route. (c)highway route. driving cycles with road slope are constructed. The average driving
power with road slope are 62.49kW (urban), 56.07kW (suburban) and
104.02kW (highway), respectively. The relative errors of average
Table 4 driving power between the real routes and representative cycles are kept
Vehicle parameters. less than 4%, i.e. -3.82% (urban), -1.01% (suburban) and -2.91%
(highway), respectively. This result shows the effectiveness of the
δ Rotating mass coefficient 1.2
representative driving cycles with road slope in terms of vehicle power.
m Mass 18 ton
If we ignore the road slope, the average driving power on the real
CD Drag coefficient 0.55
routes are 52.27 kW (urban), 49.81 kW (suburban) and 83.50 kW
A Frontal area 8 m2
(highway). The relative errors of average driving power causing by
ρ Density of air 1.2258 kg/m3
ignoring road slope become much higher than above, 24.30% (urban),
g Gravity acceleration 9.8 m/s2
13.73% (suburban) and 28.32% (highway). This result shows the ne­
f Rolling friction coefficient 0.0085
cessity of considering road slope in constructing driving cycles.
Power distribution also proves that the road slope cannot be ignored.
As shown in Fig. 11, power distribution is more accordance with the

7
X. Jia et al. Energy and AI 6 (2021) 100115

Fig 11. Power distribution.

necessary to develop a simple yet effective method for constructing

Table 6 representative driving cycles with road slope. In this paper, a 2D Markov
Errors of power distribution compared with the actual condition.
Chain method together with an average speed-based matching algo­
Urban Route Suburban Route Highway Route rithm are utilized for constructing such driving cycles in the area around
Driving cycle without slope 4.07% 5.30% 3.78% Beijing and Zhangjiakou, where plenty of fuel cell vehicles are demon­
Driving cycle with slope 2.53% 2.16% 2.13% strated now and will be developed in the future.

(1) Since road slope changes significantly in long driving mileages

actual situation after adding slope information to driving cycles, espe­
for HDVs and devices for measuring road slope are expensive, a
cially under medium to high load.
simple yet effective road slope observation technology is proposed
The difference between the power distribution under driving cycles
based on GPS signals and a Butterworth filter with an order number
and the actual condition can be measured by calculating the power
of 14, normalized pass and stop band cut-off frequencies of 0.1 and
distribution error. The power distribution error is defined as follows:
0.3. Compared with the road slope measured using the device gy­
1∑ n roscope RT3000, the observed road slope can approximate the actual
ep = |freal (Pi ) − fdc (Pi )| (9) road slope.
n i=0
(2) A 2D Markov Chain method together with an average speed-
Where, Pi refers to the power range (50i, 50i+50] kW. freal (Pi ) is the based matching algorithm is proposed for constructing representa­
proportion of power range Pi in the real data. fdc (Pi ) is the proportion of tive driving cycles with road slope. Compared to the traditional 3D
power range Pi in the driving cycles. The power distribution errors are Markov Chain method for speed, acceleration and road slope, this
shown in the Table 6. Obviously, on different routes, adding road slope method only uses speed and acceleration as two states. The road
into driving cycles can effectively reduce the power distribution errors. slope cycle is spliced with road slope segments, which are chosen
based on corresponding average speed. Three representative driving
5. Conclusion cycles with road slope around Zhangjiakou city are constructed for
urban, suburban and highway routes, respectively. Statistic results
In order to promote the electrification progress of HDVs, it is show that, the average driving power of the representative cycles are

8
X. Jia et al. Energy and AI 6 (2021) 100115

less than 4% compared to real driving conditions, indicating the [11] Lee T-K, Filipi ZS. Synthesis of real-world driving cycles using stochastic process
and statistical methodology. Int J Vehicle Des 2011;57(1):17–36.
effectiveness of this method.
[12] Shen P, Zhao Z, Li J, Zhan X. Development of a typical driving cycle for an intra-
(3) Comparison is made between average driving power of three real city hybrid electric bus with a fixed route. Transport Res D-Tr E 2018;59:346–60.
routes with and without considering of road slope. Results show that, [13] Ho S, Wong Y, Chang VW. Developing Singapore Driving Cycle for passenger cars
relative errors of average driving power caused by ignoring road to estimate fuel consumption and vehicular emissions. Atmos Environ 2014;97:
353–62.
slope are between 10~30%, indicating the necessity of considering [14] Kaymaz H, Korkmaz H, Erdal H. Development of a driving cycle for Istanbul bus
road slope in constructing driving cycles. rapid transit based on real-world data using stratified sampling method. Transport
Res D-Tr E 2019;75:123–35.
[15] Huzayyin OA, Salem H, Hassan MA. A representative urban driving cycle for
In the future, the proposed method will be utilized to design elec­ passenger vehicles to estimate fuel consumption and emission rates under real-
trified HDVs. The three representative driving cycles with road slope world driving conditions. Urban Climate 2021;36:100810.
will be used for development of hydrogen fuel cell buses and trucks. Fuel [16] Zhang L, Huang Z, Yu F, Liao S, Luo H, Zhong Z, et al. Road type-based driving
cycle development and application to estimate vehicle emissions for passenger cars
cell vehicles have been put into demonstration operation in northern in Guangzhou. Atmos Pollu Res 2021;12(8):101138.
China. But the durability and the cost remain key constraints to the [17] Cui Y, Xu H, Zou F, Chen Z, Gong K. Optimization based method to develop
further commercialization. Based on the constructed driving cycles, we representative driving cycle for real-world fuel consumption estimation. Energy
2021;235:121434.
will optimize the control strategies and energy management method to [18] Hung WT, Tong H, Lee C, Ha K, Pao L. Development of a practical driving cycle
ensure that fuel cells operate at optimal operating points. In this way, construction methodology: A case study in Hong Kong. Transport Res D-Tr E 2007;
the cost of fuel cells can be reduced, and the lifespan can be prolonged, 12(2):115–28.
[19] Barlow TJ, Latham S, Mccrae IS, Boulter PG. A reference book of driving cycles for
thus facilitating the large-scale applications..
use in the measurement of road vehicle emissions, Version 3. Bracknell, U.K: IHS;
2009.
Declaration of Competing Interest [20] Yang Y, Li T, Hu H, Zhang T, Cai X, Chen S, Qiao F. Development and emissions
performance analysis of local driving cycle for small-sized passenger cars in
Nanjing, China. Atmos Pollut Res 2019;10(5):1514–23.
The authors declare that they have no known competing financial [21] Borlaug B, Holden J, Wood E, Lee B, Fink J, Agnew S, et al. Estimating region-
interests or personal relationships that could have appeared to influence specific fuel economy in the United States from real-world driving cycles.
the work reported in this paper. Transport Res D-Tr E 2020;86:102448.
[22] Qin X, Yu K, Li H, Dai F, Liu H, Yang H, et al. Development of a one-day driving
cycle for electric ride-hailing vehicles. Transport Res D-Tr E 2020;89:102597.
Acknowledgment [23] Zhang J, Wang Z, Liu P, Zhang Z, Li X, Qu C. Driving cycles construction for electric
vehicles considering road environment: A case study in Beijing. Appl Energ 2019;
253.
This work was supported by Toyota Motor Corporation (TMC) in the [24] L. Liu, C. Huang, B. Lu, S. Shi, Y. Zhang, and J. Cheng, “Study on the design method
Tsinghua-Toyota Joint Research Center for Hydrogen Energy and Fuel of time-variant driving cycles for EV based on Markov Process,” in IEEE VPPC,
Cell Technology of Vehicles (TTFC-2019-0), and National Natural Sci­ Seoul, South Korea, pp. 1277-1281, 2012.
[25] Kolmanovsky I, Siverguina I, Lygoe B. Optimization of powertrain operating policy
ence Foundation of China (Nos. 52022050 and 52002210). for feasibility assessment and calibration: Stochastic dynamic programming
approach. P Amer Contr Conf 2002:1425–30.
Reference [26] Yang Y, Zhang Q, Wang Z, Chen Z, Cai X. Markov chain-based approach of the
driving cycle development for electric vehicle application. Enrgy Proced 2018;152:
502–7.
[1] Trimm DL, Önsan ZI. Onboard fuel conversion for hydrogen-fuel-cell-driven
[27] Wang Z, Zhang J, Liu P, Qu C, Xiaoyu L. Driving cycle construction for electric
vehicles. Cataly Rev 2001;43(1-2):31–84.
vehicles based on Markov chain and Monte Carlo method: a case study in Beijing.
[2] Ingvarsson AF, Pestiaux J, Vanek FM. A global assessment of hydrogen for future
Enrgy Proced 2019.
automotive transportation: Projected energy requirements and CO2 emissions. Int J
[28] Zhao X, Zhao X, Yu Q, Ye Y, Yu M. Development of a representative urban driving
Sustain Transp 2011;5(2):71–90.
cycle construction methodology for electric vehicles: A case study in Xi’an.
[3] Peng H, Chen Z, Deng K, Dirkes S, Ünlübayir C, Thul A, et al. A Comparison of
Transport Res D-Tr E 2020;81.
Various Universally Applicable Power Distribution Strategies for Fuel Cell Hybrid
[29] Jiang S, Wang C, Zhang C, Bai H, Xu L. Adaptive estimation of road slope and
Trains Utilizing Component Modeling at Different Levels of Detail: From
vehicle mass of fuel cell vehicle. eTransportation 2019;2:100023.
Simulation to Test Bench Measurement. eTransportation 2021:100120.
[30] B. Yue, S. Shi, N. Lin, P. Guo, Z. Li, and Z. Zhang, “Study on the design method of
[4] Hu Z, Xu L, Li J, Gan Q, Xu X, Song Z, Shao Y, Ouyang M. A novel diagnostic
driving cycle with road grade based on Markov chain model,” in IEEE VPPC,
methodology for fuel cell stack health: Performance, consistency and uniformity.
Montreal, Canada 2015, pp. 1-5.
Energ Convers Manage 2019;185:611–21.
[31] Silvas E, Hereijgers K, Peng H, Hofman T, Steinbuch M. Synthesis of Realistic
[5] Gao W. Performance comparison of a fuel cell-battery hybrid powertrain and a fuel
Driving Cycles With High Accuracy and Computational Speed, Including Slope
cell-ultracapacitor hybrid powertrain. IEEE T Veh Technol May 2005;54(3):
Information. IEEE T Veh Technol 2016;65(6):4118–28.
846–55.
[32] Souffran G, Miegeville L, Guerin P. Simulation of Real-World Vehicle Missions
[6] Qi Y, Espinoza-Andaluz M, Thern M, Andersson M. Polymer electrolyte fuel cell
Using a Stochastic Markov Model for Optimal Powertrain Sizing. IEEE T Veh
system level modelling and simulation of transient behavior. eTransportation
Technol 2012;61(8):3454–65.
2019;2:100030.
[33] China automotive test cycle- Part 2: Heavy-duty commercial vehicles, GB/T
[7] Shao Y, Xu L, Zhao X, Li J, Hu Z, Fang C, Hu J, Guo D, Ouyang M. Comparison of
38146.2, 2019.
self-humidification effect on polymer electrolyte membrane fuel cell with anodic
[34] Gagniuc PA. Markov chains: from theory to implementation and experimentation.
and cathodic exhaust gas recirculation. Int J Hydrogen Energy 2020;45(4):
NJ, USA: John Wiley & Sons; 2017. p. 1–235.
3108–22.
[35] Asmussen S. Markov chains. Applied Probability and Queues. Berlin, Germany:
[8] Tanaka S, Nagumo K, Yamamoto M, Chiba H, Yoshida K, Okano R. Fuel cell system
Springer Science & Business Media; 2003. p. 3–38.
for Honda CLARITY fuel cell. eTransportation 2020;3:100046.
[36] S. Shi, N. Lin, Y. Zhang, C. Huang, L. Liu, B. Lu, and J. Cheng, “Research on Markov
[9] Jiang H, Xu L, Li J, Hu Z, Ouyang M. Energy management and component sizing for
property analysis of driving cycle,” in IEEE VPPC, Beijing, China, 2013, pp. 453-
a fuel cell/battery/supercapacitor hybrid powertrain based on two-dimensional
457.
optimization algorithms. Energy 2019;177:386–96.
[37] Bianchi G, Sorrentino R. Electronic filter simulation & design. NewYork, NY, USA:
[10] Lin J, Niemeier DA. An exploratory analysis comparing a stochastic driving cycle to
McGraw Hill Professional; 2007. p. 17–20.
California’s regulatory cycle. Atmos Environ 2002;36(38):5759–70.