Driving Cycle Development for Urban Bus using Principal

Component Analysis and DBSCAN Clustering: With the Case of

Haikou, China

Zhenzheng Yan

1, a

, Jihui Zhuang

1, b, *

, Xiaoming Cheng

1, c

and Ying Yan

2, d

School of Mechanical and Electrical Engineering, Hainan University, Haikou 992753, China

National Local Joint Engineering Research Center for Intelligent Vehicle-Road Coordination and Safety Technology,

Tianjin University of Technology and Education, Tianjin 300222, China.

Keywords: Driving cycle, cluster analysis, PCA, DBSCAN cluster, city bus.

Abstract: Driving cycles are an important means for new vehicle technology development and emission prediction

and evaluation. To establish a representative driving cycle for urban buses in Haikou city, in this paper, the

principal component analysis (PCA) and DBSCAN cluster algorithm are applied to develop the driving

cycle. Firstly, a large number of vehicle driving data are collected, which comprised of 12 characteristic

parameters. Next, the PCA is employed to extract main components from the characteristic parameters of

driving data and the DBSCAN cluster is used to select representative micro trips. Subsequently, several

most representative micro-trips were picked out to form the driving cycle. The effectiveness and uniqueness

of the developed driving cycle are verified via comparing the parameters with the real-world driving data

and the existing driving cycles, respectively.

1 INTRODUCTION

With the increase of car ownership, vehicle

emissions have become one of the important sources

of environmental pollution (X. Han, L. P. Naeher,

2006). The main purpose of a driving cycle is to

determine vehicle pollutant emissions and fuel

consumption in the test area (Ho, Sze-Hwee, Yiik-

Diew Wong, 2014). The construction of vehicle

driving cycle is based on the investigation of the

actual driving conditions of vehicles in the test area.

Based on the statistical theory, the test data collected

on typical roads are processed and analyzed, and the

representative driving conditions of the tested area

are constructed.

In recent decades, there have been many studies in

the world on driving cycles. Europe, the United

States and Japan have constructed three world-

famous working conditions (ECE15+EUDC, FTP75

and JPAN10) according to the actual traffic

conditions of various countries. Fotouhi A (A.

Fotouhi, M. Montazeri-Gh, 2013) established

Tehran's driving cycle by K-means clustering. Qin

(D. T. Qin, S. Zhan and Z. G. QI et al, 2016)

developed driving cycles via analyzing three typical

parameters of cycle block. Dai (Dai, Zhen, et.al,

2008) and Ma (Zhixiao Ma, et.al, 2005) constructed

driving cycles using Markov stochastic method and

Dynamic Clustering Method, respectively. Each

driving cycle is unique due to different traffic and

driving conditions, data collection and analysis

technique, and vehicle type considered in the study.

On the basis of the analysis above, considering

the PCA and DBSCAN cluster can be easily

implemented due to their rigorous research

mentality, in this paper, these two methods are

combined to apply to develop the driving cycle with

case of Haikou, China. The results show that the

DBSCAN clustering algorithm make the micro-trips

clearly classified, but also reflect the traffic situation

of Haikou City.

2 DRIVING CYCLE

DEVELOPMENT

2.1 Data Collection

In order to establish a representative driving cycle of

Haikou city buses truly and effectively. We did a

108

Yan, Z., Zhuang, J., Cheng, X. and Yan, Y.

Driving Cycle Development for Urban Bus using Principal Component Analysis and DBSCAN Clustering: With the Case of Haikou, China.

DOI: 10.5220/0008872201080113

In Proceedings of 5th International Conference on Vehicle, Mechanical and Electrical Engineering (ICVMEE 2019), pages 108-113

ISBN: 978-989-758-412-1

Table 1. Representative bus routes of Haikou.

Bus routes

Route length(km) Number of bus stops Covered road types

NO.1 bus

21.4 43 Trunk roads, sub-trunk roads

NO.5 bus

10.8

Sub-trunk roads

NO.16 bus

15.7

Highways, sub-trunk roads

NO.19 bus 16.2 31 Trunk roads, sub-trunk roads

NO.21 bus

30.6 35

Highways, Trunk roads, sub-

trunk roads

NO.43 bus 23.5 44 Trunk roads, sub-trunk roads

NO.59 bus 28.2 43 Trunk roads

survey about road type and traffic situation in

Haikou City. Seven representative bus routes were

selected (see Table 1). It can be seen from Table 1,

bus routes covered a wide range of areas, including

highways, urban trunk roads, sub-trunk roads, etc.

Then, the test data of buses are collected by the

cyclic route method (Q. H. Li, 2014). Moreover, in

order to consider the effect of the travel time and

traffic flow on the driving cycle, there is 15-day test

was conducted. The test period is from 6:30 a.m. to

22:30 p.m., and the test data included off-peak hours

and peak hours, covering non-working days and

working days.

The research data in this study came from the

On-board data acquisition terminal specified in this

project, as portrayed in Figure 1. The information of

city buses’ position, speed and acceleration can be

obtained from the terminal, and it will be transmitted

to the data monitoring platform through the

4G network for later data analysis and

processing. Figure 2 presents a flow chart of the

road test remote information system.

Fig 1. On-board data acquisition terminal.

2.2 Data Preprocessing

Vehicles may be influenced by various traffic

conditions that result in several start–stop operations

(Y. B. Zheng, 2014) throughout the process. In this

paper, the motion of a vehicle from one idle to the

next idle is defined as a micro-trip (S.H. Kamble,

et.al, 2009). The collected data were divided into

3423 micro-trips according to the Table2. To

facilitate the classification of 3423 micro-trips, the

selections of assessment criteria are picked out from

the following characteristic parameters in Table 3.

Table 2. Driving modes of micro-trips.

driving modes Description

Accelerate mode a≥0.15m/s2, v≠0

Cruise mode |a|≤0.15m/s2, v≠0

Deceleration mode a≤-0.15m/s2 , v≠0

Idle mode Engine working, v=0, a=0

2.3 Methodology

2.3.1 Principal Component Analysis (PCA)

Visually finding the pattern and the law is difficult

with high-dimensional data. Therefore, dimension

reduction was necessary in this study. Different

strength correlations were noted among of the 12

characteristic parameters used for classification, and

they were not independent of one another. The

principal component analysis (PCA) is a method for

reducing the size of a given collection of data while

keeping the information of the original data (Z. Jing,

et.al, 2017). Hence, in this paper, PCA was used to

reduce the dimension of these parameters first. The

PCA of 12 parameters was carried out by SPSS, and

the characteristic values and contribution rates of

each principal component were obtained, as shown

in Table 4.

Driving Cycle Development for Urban Bus using Principal Component Analysis and DBSCAN Clustering: With the Case of Haikou, China

109

Fig 2. Data collection process.

Table 3. Characteristic parameters.

parameters Implication

Vm average speed (km/h)

Vmax maximum speed(km/h)

Vsd Standard deviation of

velocity(km/h)

Amax maximum acceleration(m/s

)

Amin minimum deceleration(m/s

)

Asd Standard deviation of

acceleration(m/s

)

Pa acceleration time ratio

Pd deceleration time ratio

Pi idle time ratio

Pcon constant time ratio

Am average acceleration(m/s

)

Amd average deceleration(m/s

)

As exemplified in Table 4, the cumulative

contribution rate of former 4 principal components

to total variation accounted for over 90%. It

indicates the first 4 principal components maintain

most of information of 12 characteristic parameters

mentioned earlier. Namely every micro-trip can be

described by 4 principal components instead of 12

characteristic parameters for further analysis. In

other word, the dimension of sample data is reduced

from 3423*12 to 3423*4, which can reduce the

difficulty of the analysis.

2.3.2 DBSCAN Clustering

DBSCAN (Density-Based Spatial Clustering of

Application with Noise) clustering algorithm is a

method of density-based unsupervised clustering

algorithm (Lizhao Han, et.al, 2018). Compared with

traditional K-Means algorithm, DBSCAN has the

advantage of clustering dense data sets of arbitrary

shape. At the same time, it can also find noise points

while clustering. It is undeniable that the clustering

effect of this algorithm is not ideal when facing

high-dimensional data sets. However, combined

with the role of PCA dimensionality reduction, we

can make up for this defect and get accurate

clustering results.

Two key parameters of DBSCAN algorithm

should be determined: one is the radius (Eps) which

represent the range of circular neighborhoods

centered on a given point P; the other is the number

of minimal points (MinPts) in the neighborhood

centered on point P. The flow of DBSCAN

algorithm is concluded in Figure 3.

vehicle

CAN

network

Data

monitoring

platform

engine speed

vehicle speed

accelerator pedal

depth

intake air flow

engine torque

GPS time

GPS location

network

terminal

ICVMEE 2019 - 5th International Conference on Vehicle, Mechanical and Electrical Engineering

110

Table 4. Result of PCA.

Componen

Total Variance% Cumulative%

1 6.609 44.063 44.063

2 3.710 24.734 68.797

3 1.858 12.389 81.186

4 0.610 4.068 92.709

5 0.358 2.385 95.094

6 0.345 2.299 97.393

7 0.179 1.194 98.587

8 0.116 0.773 99.360

9 0.023 0.154 99.739

10 0.013 0.087 99.935

11 0.010 0.065 100.000

12 5.59E-9 3.72E-8 100.000

Fig 3. Process of DBSCAN clustering algorithm.

2.4 Composition of Driving Cycle

On the basis of the analysis above, DBSCAN

algorithm can be used to cluster sample data after

processed by PCA. Sample data was separated into

two classes by this algorithm: including 2471 micro-

trips belong to the first class while 952 micro-trips

were classified as the second class. According to the

time ratio of the two categories and the correlation

coefficient of candidate fragments (Q. Shi, D. Y.

Qiu and J. Y. Zhou, 2012), 11 micro-trips were

selected from the first category and 5 micro-trips

were selected from the second category, as shown in

Table 5. The representative driving cycle of Haikou

city buses was built by combining these selected

micro-trips (see Figure 4).



Fig 4. Driving cycle of Haikou city buses.

3 RESULTS AND DISCUSSION

3.1 Parametric Analysis

To confirm the efficiency of driving cycle as the test

cycle, in this paper, compare the driving cycle with

the 12 characteristic parameters of the real-world

driving data. The results are displayed in Table 6. It

can be observed in Table 6 that the average error of

these two sets of figures is 5.8%, which indicates the

Haikou driving cycle can reflect the real-world

traffic conditions.

Start

Input data set X

Set up Eps and MinPts

Select an unvisited point P

Joint Core points

yes

Find all border points whose

pn<MinPts but in P's Esp

Output clustering results

End

PN>MinPts

Calculate PN of P

neighborhood with Esp

Marked

as noise

point

Unvisite d points in X

Joint border points to core points

yes

Driving Cycle Development for Urban Bus using Principal Component Analysis and DBSCAN Clustering: With the Case of Haikou, China

111

Table 5. Driving parameters of 16 micro-trips.

Fragments 1 2 3 ... 16

Duration(s) 154 50 58 ... 54

Acceleration(s) 43 17 21 ... 13

Constant velocity(s) 75 7 11 ... 12

Deceleration(s) 27 21 11 ... 16

Idle time(s) 9 5 15 ... 13

Vmax (km/h) 58.61 42.34 39.44 ... 53

Vm (km/h) 29.35 22.18 19.41 ... 23.15

Vsd (km/h) 19.87 15.31 15.86 ... 16.12

Amax (m/s2) 1.32 1.14 1.23 ... 2.15

Am (m/s2) 0.40 0.68 0.51 ... 0.67

Amin (m/s2) -1.47 -1.18 -2.09 ... -2.08

Amd (m/s2) -0.59 -0.51 -0.90 ... -0.51

Asd (m/s2) 0.21 0.24 0.28 ... 0.27

3.2 Contrast with Typical City Driving

Cycles

To further evaluate the developed Haikou driving

cycle, comparison with the existing standard driving

cycles namely ECE15, FTP75 and JPAN10. The

comparison is based on key parameters as illustrated

in Table 7. It can be observed that, the driving cycle

of Haikou significantly differed from typical driving

cycles in terms of most parameters. Generally

speaking, the buses in Haikou city, compared with

other driving cycles, were in a low-speed state with

frequent acceleration and deceleration.

Table 6. Parametric analysis between two conditions.

Parameters

Real-

world

driving

data

Haikou

driving

cycle

Relative

error

(%)

Vm 24.13 22.43 7.04

Vmax 70.30 63.12 10.21

Vsd 12.21 13.25 8.51

Amax 2.89 2.75 4.84

Amin -2.78 -2.62 5.76

Asd 0.30 0.26 13.3

Pa(%) 30.35 31.17 2.7

Pd(%) 24.83 25.19 1.45

Pi(%) 18.20 17.23 5.33

Pcon(%) 26.62 26.41 0.79

Am 0.547 0.592 2.74

Amd -0.603 -0.537 4.31

4 CONCLUSIONS

In this paper, a methodology for development of

driving cycle using PCA and DBSCAN algorithm is

proposed. The driving cycle for the city of Haikou is

developed by this method and it is validated a highly

representative compared with the real-world driving

data. Besides, the developed driving cycle was

compared with the existing driving cycles, thus

highlighting the uniqueness of the traffic conditions

in Haikou. Namely, the buses in Haikou City are

maintained at low-speed for most of the time and

have frequent acceleration and deceleration actions.

Table 7. Contrast with typical city driving cycles.

Parameters

Haikou driving

cycle

ECE15 FTP75 JPAN10

Vm 22.43 34.1 18.4 17.6

Amax 2.75 1.44 0.81 0.81

Amin -2.62 -1.44 -0.81 -0.81

Pa(%) 31.17 32.4 21.5 24.3

Pd(%) 25.19 28.2 18.5 25.0

Pi(%) 17.23 17.9 30.8 27.2

Pcon(%) 26.41 21.5 29.2 23.5

Am 0.59 0.61 0.64 0.67

Amd -0.54 −0.7 −0.75 −0.65

ICVMEE 2019 - 5th International Conference on Vehicle, Mechanical and Electrical Engineering

112

REFERENCES

X. Han, L. P. Naeher, “A review of traffic-related air

pollution exposure assessment studies in the

developing world,” Environment international, Vol. 1

(2006), No. 32, p. 106-120.

Ho, Sze-Hwee, Yiik-Diew Wong, and Victor Wei-Chung

Chang. "Developing Singapore Driving Cycle for

passenger cars to estimate fuel consumption and

vehicular emissions." Atmospheric environment 97

(2014), p. 353-362.

A. Fotouhi, M. Montazeri-Gh, “Tehran driving cycle

development using the k-means clustering method,”

Scientia Iranica, Vol. 2 (2013), No. 20, p. 286-293.

D. T. Qin, S. Zhan and Z. G. QI et al., “Driving cycle

construction using K-means clustering method,”

Journal of Jilin University (Engineering and

Technology Edition), Vol. 2 (2016), No. 46, p. 383-

389.

Dai, Zhen, Deb Niemeier, and Douglas Eisinger. "Driving

cycles: a new cycle-building method that better

represents real-world emissions." Department of Civil

and Environmental Engineering, University of

California, Davis (2008).

Zhixiao Ma, Xichan Zhu, Mengliang Li, et al.

"Application of Dynamic Cluster to the Development

of Vehicle Real World Driving Cycle."Journal of

Wuhan University of Technology, Vol. 11 (2005), No.

27,p. 69-71.

Q. H. Li, “Analysis of Driving cycle of Road Vehicles in

Hefei,” Technology & Economy in Areas of

Communications, Vol. 3 (2014), p.100-103.

Y. B. Zheng, Q. Shi and S. L. Wang, “Research of

Automobile Driving Cycle in Hefei City,” Automobile

Technology, Vol. 10 (2014), p. 34-39.

S.H. Kamble, T.V. Mathew and G.K. Sharma,

“Development of real-world driving cycle: case study

of Pune, India,” Transport. Res. Part D Transp.

Environ, Vol. 2, no. 14 (2009), p.132-140.

Z. Jing, G. Wang and S. Zhang et al., “Building Tianjin

driving cycle based on linear discriminant analysis,”

Transportation Research Part D: Transport and

Environment, vol. 53(2017), p. 78-87.

Lizhao Han, Xuezhong Qian, Jing Luo, et al., “Multi-

density clustering algorithm DBSCAN based on

region division.” Application Research of Computers,

Vol. 35 (2018), No. 6, p.74-77, 91.

Q. Shi, D. Y. Qiu and J. Y. Zhou, “Driving Cycle

Construction and Accuracy Analysis Based on

Combined Clustering Technique,” Automotive

Engineering, vol. 2 (2012), no. 34, p. 164-169, 158.

Driving Cycle Development for Urban Bus using Principal Component Analysis and DBSCAN Clustering: With the Case of Haikou, China

113