ArticlePDF Available

Impact of Driver Behavior on Fuel Consumption: Classification, Evaluation and Prediction Using Machine Learning

June 2019
IEEE Access PP(99):1-1

DOI:10.1109/ACCESS.2019.2920489

License
CC BY-NC-ND 4.0

Authors:

Ping Peng

Southeast University (China)

Yang Xu

Nanjing University of Information Science & Technology

Show all 5 authorsHide

Driving behavior has a large impact on vehicle fuel consumption. Dedicated study on relationship between driving behavior and fuel consumption can contribute to decrease the energy cost of transportation and the development of the behavior assessment technology for the ADAS system. So, it is vital to evaluate this relationship in order to develop more ecological driving assistance systems and improve vehicle fuel economy. However, modeling driving behavior under dynamic driving conditions is complex, making quantitative analysis of the relationship between driving behavior and fuel consumption difficult. In this paper, we introduce two kinds of machine learning methods for evaluating the fuel efficiency of driving behavior using naturalistic driving data. In the first stage, we use an unsupervised spectral clustering algorithm to study the macroscopic relationship between driving behavior and fuel consumption, using data collected during the natural driving process. In the second stage, dynamic information from the driving environment and natural driving data are integrated to generate a model of the relationship between various driving behaviors and the corresponding fuel consumption features. The dynamic environment factors are coded into a processible, digital form using a deep learning-based object detection method, so that the environmental data can be linked with the vehicle’s operating signal data to provide the training data for the deep learning network. The training data is labeled according to its fuel consumption feature distribution, which is obtained from road segment data and historical driving data. This deep learning-based model can then be used as a predictor of the fuel consumption associated with different driving behaviors. Our results show that the proposed method can effectively identify the relationship between driving behavior and fuel consumption on both macro and micro levels, allowing for end-to-end fuel consumption feature prediction, which can then be applied in advanced driving assistance systems.

Two-stage architecture of the proposed driving behavior modeling method. In the first stage (outlined in red) unsupervised machine learning is used to obtain the macro-level fuel consumption features of driver behavior. In the second stage (outlined in blue) an LSTM is used to analyze short-term driving behavior and driving environment data to predict real-time fuel consumption.

…

Two types of roads used for data collection. Left: expressway loop with two lanes. Right: ordinary road with one lane in each direction.

…

Overall map of the roads used for data collection. Yellow line is the expressway and the white line is the ordinary road.

…

Age and sex distribution of all the experiment participants.

…

+22

Data collection system (for driving data and GPS information).

…

Figures - uploaded by Ping Peng

Content may be subject to copyright.

Content uploaded by Ping Peng

Content may be subject to copyright.

Available via license: CC BY-NC-ND 4.0

Content may be subject to copyright.

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.

Digital Object Identiﬁer 10.1109/ACCESS.2017.DOI

Impact of driver behavior on fuel

consumption: classiﬁcation, evaluation

and prediction using machine learning

PENG PING1, WENHU QIN1, YANG XU1, CHIYOMI MIYAJIMA2,(Member, IEEE) and KAZUYA

TAKEDA3, (Senior Member, IEEE).

1School of Instrument Science and Engineering, Southeast University, Nanjing, 210096, China

2School of Informatics, Daido University, Nagoya, 457-8530, Japan

3The Graduate School of Informatics, Nagoya University, Nagoya, 464-0814, Japan

Corresponding author: Wenhu Qin (e-mail: qinwenhu@seu.edu.cn).

ABSTRACT Driving behavior has a large impact on vehicle fuel consumption. Dedicated study on

relationship between driving behavior and fuel consumption can contribute to decrease the energy cost

of transportation and the development of the behavior assessment technology for the ADAS system. So, it is

vital to evaluate this relationship in order to develop more ecological driving assistance systems and improve

vehicle fuel economy. However, modeling driving behavior under dynamic driving conditions is complex,

making quantitative analysis of the relationship between driving behavior and fuel consumption difﬁcult.

In this paper, we introduce two kinds of machine learning methods for evaluating the fuel efﬁciency of

driving behavior using naturalistic driving data. In the ﬁrst stage, we use an unsupervised spectral clustering

algorithm to study the macroscopic relationship between driving behavior and fuel consumption, using data

collected during the natural driving process. In the second stage, dynamic information from the driving

environment and natural driving data are integrated to generate a model of the relationship between various

driving behaviors and the corresponding fuel consumption features. The dynamic environment factors are

coded into a processible, digital form using a deep learning-based object detection method, so that the

environmental data can be linked with the vehicle’s operating signal data to provide the training data for the

deep learning network. The training data is labeled according to its fuel consumption feature distribution,

which is obtained from road segment data and historical driving data. This deep learning-based model can

then be used as a predictor of the fuel consumption associated with different driving behaviors. Our results

show that the proposed method can effectively identify the relationship between driving behavior and fuel

consumption on both macro and micro levels, allowing for end-to-end fuel consumption feature prediction,

which can then be applied in advanced driving assistance systems.

INDEX TERMS Driving behavior modeling, Data mining, Deep learning, Vehicle fuel economy.

I. INTRODUCTION

ACombination of emissions from coal combustion and

urban vehicle use has become the primary source of air

pollution in most of the world’s major cities [1,2]. According

to the World Health Organization, transportation emissions

are a signiﬁcant and growing contributor to particulate air

pollution, which makes up 30% of particulate matter emis-

sions (PM) in European cities and 50% of PM emissions

in OECD countries [3]. One study estimated that approxi-

mately 1.03 million deaths were associated with ambient PM

2.5 air pollution in the 74 largest cities of China in 2013,

which accounted for 32% of all reported deaths [4]. As a

result, much research has been focused on reducing vehicle

emissions. As has been demonstrated in various studies [5-

7], driving behavior, such as speed control, preferred rate

of acceleration, and vehicle control stability, have a major

effect on fuel consumption, regardless of the type of vehicle

being driven. By accurately identifying relationships between

driving behavior and fuel consumption, Advanced Driving

Assistant Systems (ADAS) can be designed to give more

accurate and intelligent eco-driving advice [8,9]. By studying

the driving behavior’s impact on the fuel consumption, we

VOLUME 4, 2016 1

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

can know how some drivers cost more energy than others

so as to help high energy cost drivers to achieve fuel-

effect driving style. Besides, as the fundamental technology

of the ADAS systems or eco-driving coaching system, the

effective driving behavior-energy consumption model can

be applied to decrease the commercial vehicle’s fuel cost

[10], optimal the charging station location [11], decrease

the transportation’s greenhouse gas emission [12] and so on.

Thus, discovering the precise relationship between driving

behavior and fuel consumption, in order to reduce vehicle

emissions and increase fuel efﬁciency, has become an impor-

tant studying area and the motivation of our study. However,

effective analysis model for driving behavior’s impact on fuel

consumption is rarely studied. In this paper, we aimed to

design a machine learning based method which can analyse

and predict a reasonable relationship between the driving

behavior and fuel consumption. The eco-driving system or

ADAS system can obtain driving state from the proposed

model so as to give more reasonable advice to the driver to

keep fuel-efﬁcient driving.

Quantitative analysis of the relationship between driving

behavior and fuel consumption is a natural and direct ap-

proach. However, traditional fuel consumption models such

as the Vehicle Speciﬁc Power (VSP) model [13], the Com-

prehensive Modal Emission Model (CMEM) [14] and the

International Vehicle Emissions model (IVE) [15] are specif-

ically designed to evaluate the fuel economy performance of

engines, and the process of calibrating these models is very

complex [16]. In contrast, most driving behavior modeling

studies have focused on speciﬁc driving scenarios, such as

lane changes [17,18], arterial corridors [19], signalized inter-

sections [20], and so on. These models focus on identifying

safe or comfortable driving, which are difﬁcult to link to fuel

consumption. As a result, the integration of driving behav-

ior parameters or models with traditional fuel consumption

models is a problem which remains to be resolved. Many re-

searchers have proposed two-stage methods, where statistical

or machine learning methods are used to identify a driver’s

driving style, and then the features of that driving style

are compared with the related fuel consumption features. J.

E. Meseguer et al. used a three-layered neural network to

classify drivers into quiet, normal and aggressive groups [21].

They then analyzed the fuel consumption features for each

group. E. Gilman et al. used 17 driving behavior factors to

identify fuel-efﬁcient driving behavior for a driver coaching

system [22]. The driving behavior factors were evaluated

according to their distributions, calculated from a historical

driving trip. R. Trigui et al. analyzed the impact of vari-

ous driving behaviors on fuel efﬁciency using mathematical

modelling [23]. The study ﬁrst divided driving behavior into

two levels; maneuvering level and control level behavior.

Then, by identifying the various parameters of their model,

the authors simulated three different behaviors; aggressive

driving, eco-driving and normal driving. Their results showed

that their proposed model could accurately match measured

fuel consumption and real driving behavior. C. Lv et al.

proposed an unsupervised machine learning method using

Gaussian mixture models to recognize three typical driving

styles, and then provided the optimal control strategy for each

driving style in order to improve energy efﬁciency [24]. All of

the studies cited here succeeded in identifying fuel-efﬁcient

driving behavior, however their lack of detailed consideration

of the impact of various trafﬁc condition limits the usefulness

of their results as driving behaviors are also inﬂuenced by

various static or dynamic environmental factors [25,26].

Therefore, some researchers have also examined driving

environment features, which can be deduced or directly

obtained from naturalistic driving data, in their analyses of

driver fuel consumption, resulting in more nuanced assess-

ments. M. Ehsani et al. discussed in detail the effects of

external environmental factors on vehicle fuel consumption

[27], but did not carefully examine the effect of driving

behavior, only mentioning that speed and acceleration are

the two most important parameters. J. Rios-Torres et al.

classiﬁed driving styles into three categories by analyzing

real-world data, and then examined the effect of each driving

style on fuel consumption [28]. The results of this study show

that vehicle fuel consumption can vary widely compared with

standard US Environmental Protection Agency (EPA) driving

cycles, depending on the driver’s driving style and the driving

scenario.

The studies mentioned above investigating the relation-

ship between driving behavior and fuel consumption have

achieved good results, but unanswered questions remain.

Most of these studies have employed statistical or rule-based

methods to analyze the relationship between driving behav-

ior and fuel consumption, so these methods require huge

amounts of long-term driving data as well as prior knowledge

of the data’s statistical feature. The ordinary methods usually

need lots of expert skills to extracted prior knowledge from

the raw data set. And the results have limited universality

because the experiments have mostly been conducted on a

limited variety of trafﬁc conditions. Although the machine

learning method also need considerable amount of data, the

learning-based method can learn the inner feature or the

knowledge from the raw data automatically.

Thus, in this paper we propose an approach which employs

two machine learning methods, in order to push the research

of the fuel-efﬁcient driving behavior one step further. In

the ﬁrst stage, we use an unsupervised machine learning

method to analyze the fuel efﬁciency of driver behavior

macroscopically, as shown in the upper section of Fig. 1

(circled in red). Inspired by some previous studies [29-31]

in which machine learning was used for driving behavior

analysis, in this study we employ a parallel spectral clustering

algorithm [32] to classify the driving signal dataset collected

from multiple drivers. Drivers are divided into three groups

based on similarities in their driving styles. We then analyze

the data to extract the data points which lie in the same

fuel consumption zone. Due to the properties of spectral

clustering, prior knowledge about the data is not required, so

this clustering method is suitable for dealing with unique sets

2VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 1: Two-stage architecture of the proposed driving behavior modeling method. In the ﬁrst stage (outlined in red)

unsupervised machine learning is used to obtain the macro-level fuel consumption features of driver behavior. In the second

stage (outlined in blue) an LSTM is used to analyze short-term driving behavior and driving environment data to predict real-time

fuel consumption.

of driving data. A parallel calculating structure is also used

to improve the efﬁciency of the clustering process.

The other machine-learning method used in this study is

Long Short-Term Memory (LSTM), which is a powerful

method for modeling behavior [33]. In contrast to previous

studies which using LSTM to analyse the fuel consumption

model [34,35], in this paper we include more features of the

dynamic trafﬁc environment, in form of video frames, in our

learning model, as shown in the lower part of Fig. 1 (circled

in blue), so as to make the network more robust and general

to a wider variety of trafﬁc conditions. In addition to ana-

lyzing the fuel efﬁciency of a driver’s historic or long-term

driving behavior, our learning-based method is designed to

also examine short-term driving data, making the prediction

results adaptive to dynamic trafﬁc conditions. The input end

of the model uses video frame, GPS and ECU information,

while the output is a real-time prediction of the level of fuel

consumption. This structure allows end-to-end evaluation of

the fuel-efﬁciency of driving behavior.

This paper is organized as follows: The paper’s objectives

and related research are described in the Introduction. Section

II provides details about the spectral clustering algorithm we

employed and describes the collection of driving behavior

data using data mining. Section III describes our use of an

LSTM to predict short-term fuel consumption features and

describes the model’s performance using representative fuel

consumption feature prediction results. Finally, in Section IV

we discuss our ﬁndings and conclusions.

II. DATA COLLECTION AND UNSUPERVISED

EXTRACTION OF FEATURES OF FUEL-EFFICIENT

DRIVING BEHAVIOR

A. DATA COLLECTION PROCESS

1) Experiment design

Research by E. Ericsson [26] suggests that driving behav-

ior is affected by various factors such as street design, trafﬁc

management methods, trafﬁc conditions, weather conditions

and the driver’s mental and physical condition. In order

to evaluate the effect of the driver’s condition on vehicle

fuel consumption and simplify the veriﬁcation process, in

this study we ﬁxed the vehicle type, trip route and weather

conditions used in our experiment. The only variable factors

are the drivers (i.e., their driving behavior) and the trafﬁc con-

ditions. If more than one route were used in the experiment, it

would be difﬁcult to determine which factors were primarily

responsible for variation in fuel consumption. Therefore, all

of the data for our experiment was collected using a ﬁxed

route which included some variation in road types. Examples

of the two types of roads used in our study are shown in Fig.

2. The total distance of all of the road segments was about

VOLUME 4, 2016 3

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 2: Two types of roads used for data collection. Left:

expressway loop with two lanes. Right: ordinary road with one

lane in each direction.

15.2 km, which consisted of a 5.3 km expressway loop with

two lanes in each direction and 9.9 km of ordinary road with

one lane in each direction. The detail route map and road

information are shown in Fig. 3.

FIGURE 3: Overall map of the roads used for data collection.

Yellow line is the expressway and the white line is the ordinary

road.

Our data was collected using 30 normal passenger cars

with a 1.2T (85kw) gasoline engine and a six-speed auto-

matic transmission (6AT). Fuel consumption increases by

0.38±0.079% each time the air temperature decreases by 1◦C

[36]. Therefore, in order to avoid the possibility of variations

in air temperature obscuring the relationship between driv-

ing behavior and fuel consumption, the data collection was

conducted in the autumn from September to November. 202

drivers are selected to join the experiment, the information

of the drivers is shown in Fig. 4. As the supervised and

unsupervised learning method need lots of samples, so we

try out best to ﬁnd the experiment participants as much as

possible. We choose these 202 drivers from our university’s

students and the cooperator’s staffs. All the participants drove

in the experimental route for 10 circuits a day and the whole

experiment of single drivers last a week. When processing

our experiment, we did not give time limitation or some

special driving tasks to the participants in order to avoid

extra mental pressure. We just tell them the research goal,

experimental route and drive as they usually do. Most of the

experiment participants are in normal emotion and will be

FIGURE 4: Age and sex distribution of all the experiment

participants.

FIGURE 5: Data collection system (for driving data and GPS

information).

paid after the experiment.

2) Data collection and redundant data pruning

The data collection system (DCS) in Fig. 5 is divided

into three parts: a vehicle-mounted data collection system

(VMDCS), a wireless transmission system (WTS) and a

data center (DC). The VMDCS uses On-Board Diagnostics

(OBD) to obtain the vehicle’s operating information from

the ECU, and uses GPS to track the vehicle’s position. The

WTS uses a wireless transmission unit (WTU) installed on

the vehicle which communicates with the base station via

4G broadband to upload the collected data. Messages from

the WTS include a receiving module IP address so that the

data can be transmitted to the DC via the internet. The DC

server shows the vehicle’s position and real-time vehicle

information on the Web. The collected data is stored in an

SQL database.

In order to improve calculation efﬁciency, we selected

vehicle operation data with a strong relationship to driving

behavior, and used the Pearson correlation coefﬁcient (PCC)

4VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 6: Correlation coefﬁcients of various vehicle op-

eration parameters with fuel consumption. Red bar: strong

correlation, Yellow bar: moderate correlation, Blue bar: weak

correlation.

TABLE 1: Difference between calculated fuel consumption

and fuel consumption analyzer results.

Vehicle load

Road type Urban road Expressway Rural road

No-load 4.85% 1.28% 2.18%

Full-load 5.94% 0.81% 3.65%

[37] to determine the relevance of each parameter to vehicle

fuel consumption. We treated positive and negative accel-

eration as different parameters because their effects on fuel

economy differ. For example, when calculating fuel cost,

if negative acceleration is less than zero, instantaneous fuel

consumption is zero. The calculated correlation coefﬁcients

for various features are listed in Fig. 6, where PCC value ρis

represented by different color bars according to the following

standard guidelines; when |ρ|>0.5= strong correlation,

when 0.5>|ρ|>0.3= moderate correlation, when

|ρ|<0.3= weak correlation [38]. In Fig. 6, ‘Negative acc’

and ‘Negative acc variance’ have a negative correlation with

fuel consumption, so in fact, the PCC of these two parameters

are negative values. Then, before using an unsupervised

clustering method to abstract the data distribution features,

we ﬁrst pruned the weakly correlated data parameters.

3) Fuel consumption calculation

To calculate fuel consumption, we integrated instant fuel

consumption information from the ECU to obtain accumu-

lated fuel consumption data. In order to verify the results

of our calculations, we compared our calculated results with

the results from a fuel consumption analyzer under various

trafﬁc conditions. The differences between these two fuel

consumption measurement approaches are shown in Table 1.

From the data in Table 1, we can conclude that the dif-

ference between our calculation method and actual fuel con-

sumption is less than 6%. As the route used in our experiment

is only 15 km in length and the goal of the study is to

evaluate the effect of driving behavior on fuel consumption,

this difference can be ignored.

B. DATA SEGMENT CONSTRUCTION

As our research goal is to analyze and predict the impact

of driving behavior on fuel consumption within a limited

time frame (25 to 35 minutes), in this section we describe

the spectral clustering method we used to compare inner

similarity within the data set, so as to cluster data with similar

features into the same cluster. Our spectral clustering method

can only handle data sets of the same size. The data collection

rate was 10Hz and we collected 15,000-21,000 data points

per circuit of the driving route (we treated each circuit of the

driving route as an independent data set). Since the amount of

data collected in each data set varied, we needed to compress

each data set to the same size.

FIGURE 7: Data compression process based on road seg-

ment.

As shown in Fig. 7, we ﬁrstly partitioned the raw data

set into several subsets. The driving route was divided into

50 road segments according to their location distribution.

And then the whole data will be divided according to their

belonging road segment (each data points contain the GPS

position). As each road segment contains a different num-

ber of data points, we needed to calculate each segment’s

minimum data size Sn. For example, S1is the minimum

data size of the ﬁrst road segment (calculated from the entire

data set associated with the ﬁrst road segment). Each data set

allocated to road segment 1 is then compressed to size S1.

After data compression, each data set will have the same data

size Sall, as shown in equation (1):

Sall =

i=1

Si(1)

In contrast to using maximum information entropy to select

the size limit of the data, as in our previous study [39], the

VOLUME 4, 2016 5

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

data compression method adopted in this paper allows us to

retain most of the data points.

C. UNSUPERVISED DATA FEATURE EXTRACTION

1) Spectral Clustering Algorithm

Unsupervised machine learning is usually used for data

distribution analysis or data set inner feature abstraction. In

this paper, we adopt spectral clustering to study the features

of our self-collected dataset. As described previously, we

collected driving data sets of the same size from multiple

drivers during natural driving along a ﬁxed route. Spectral

clustering performs data clustering as a graph partitioning

problem without making any assumptions about the form of

the data clusters. Due to this characteristic, we do not need

to have prior knowledge of the driving behavior data. This is

very important for our research because the data sets which

are obtained from the data collection platform vary from

driver to driver. Spectral clustering is a suitable method for

working with these kinds of ‘random’ data sets. In additions,

spectral clustering is reasonably fast, especially for sparse

data sets of up to several thousands of points. Furthermore,

spectral clustering is not dependent on the dimensions of the

data sets. The ﬁrst step of the spectral clustering process is to

construct driving data layout graph G, which is an undirected

similarity graph for the parameters of the data points, all of

which are scalar. We use Xto represent the entire raw driving

data set:

X = {x1, x2, . . . , xN}, xi∈Rl×Sall (2)

Each xicontains the six selected fuel-efﬁciency linked

parameters which were chosen as described above, so l = 6 in

this case. N is the total number of data samples. Graph G is

weighted using the distances between each pair of vertices xi

and xj, which are represented by non-negative weight wi,j.

Because there has been no deﬁnitive determination of how

the designs of similarity graphs inﬂuence spectral clustering

results [29], here we use a full-connection to construct simi-

larity matrix W, and use a Gaussian function to calculate wi,j

as follows:

wi,j = exp −kxi−xjk2

2δ2!, δ = 10 (3)

Similarity matrix W∈RN×Nis constructed using the

terms of wi,j . Obviously, matrix W is a symmetric matrix

for G, which is an undirected similarity graph. We then build

degree matrix D, which is a diagonal matrix with degree

(d1, . . . , dn)as the diagonal. The degree of vertex xiis

deﬁned as:

di=

j=1

wi,j (4)

Two other parameters are deﬁned, the volume of a cluster,

Vol(C), and the border between two clusters, Cut (C1, C2),

which are calculated as follows:

Vol(C) = X

i∈C

di(5)

Cut (C1, C2) = X

i∈C1X

j∈C2

wi,j (6)

Next, similarity graph G is partitioned into disjointed sets.

There are different graph cutting methods, such as MinCut

[40], RatioCut [41] and NCut [42]. MinCut is simple and

effective, but it often fails to satisfactorily solve the problem

due to possible singularity problems. RatioCut and NCut

take into consideration the vertices and edge weights to

make the clusters more balanced, but RatioCut is relatively

slow, so in this study we chose Ncut, which is an NP-hard

problem [40], as our border determination method. In order to

obtain optimal clustering results, we used the object function

shown in (7), where (A1, . . . , Ak)are the ﬁnal clustering

groups. This object function is used again in (10). Aiis the

complementary set of Ai:

minNcut(A1, ..., Ak) =min(1

i=1

W(Ai, Ai)

Vol(Ai))

= min(

i=1

Cut(Ai, Ai)

Vol(Ai))

(7)

A group of indicator vectors hj= (h1,j, . . . , hn,j)Tare then

deﬁned as follows:

hi,j =(1

√Vol(Aj), xi∈Aj

0, xi/∈Aj

(8)

Matrix H∈RN×kwhich contains the k indicator vec-

tors hi,j as columns, is then constructed. Normalized graph

Laplacians [44] are then introduced as:

Lsym =D−1

2LD−1

2=I−D−1

2W D−1

2(9)

Due to the following given properties:







H0H=I

iDhi= 1

iLhi= Cut Ai, Ai/Vol (Ai)

(10)

the Ncut problem is then reformulated as:

argminA1,...AkTr (H0LH )subject to H0DH =I(11)

By substituting T = D−1

2H, we can change the Ncut

problem into a simpler form:

argminT∈RN×kTr T0D−1

2LD−1

2Tsubject to T0T=I

(12)

Then, according to the Rayleigh-Ritz theorem [32,45], this

standard trace minimization problem can be solved using

matrix U, which contains k eigenvectors as columns, corre-

sponding to the ﬁrst k eigenvalues (in increasing order) of

Lsym. Finally, by taking each row of matrix U as new data

sets, we then cluster them into k groups using a K-means

6VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

clustering algorithm. If the unit in row i of matrix U belongs

to group Cj, the original data xiin the raw data set X also

belongs to group Cj.

2) Parallel spectral clustering algorithm

The time complexity of a spectral clustering algorithm is

O(n3), where n represents the amount of data. If n is greater

than 5,000, the time cost of spectral clustering using conven-

tional calculation methods will be excessively high, therefore

we introduce a method of parallel spectral clustering which

employs cloud computing. The cloud computing platform

Spark [46] is suitable for parallel calculations involving big

data. By analyzing the inner calculation mechanism of our

spectral clustering method, we see that three processes are

responsible for most of the calculation time cost: construction

of the similarity matrix, calculation of the eigenvalues of the

graph Laplacians and the ﬁnal K-means clustering.

The process of parallel spectral clustering using the Spark

platform can be described as follows:

Step 1: Calculating the similarity matrix in a parallel

manner.

First, we store the entire raw data set in a Hadoop dis-

tributed ﬁle system (HDFS), since data sets in HDFS can

be accessed by the whole calculating cluster. We then use

the Spark resilient distributed dataset (RDD) map method

(shown in Fig. 8) to assign the spilt data set to several

parallel calculating tasks. Because the similarity graph is

fully connected, the similarity matrix is symmetrical. As a

result, we just need to calculate wi,j ,∀1≤i≤j≤N. The

detailed method for dividing the data to construct sub-sets is

shown below:

Raw data set: X = x1, x2, . . . , xN

Fragment set:

X1={x1, X0

1}, X −X0

1=∅

X1={x2, X0

1}, X −X0

1={x1}

XN={xN, X0

N}, X −X0

N={x1, x2, . . . , xN−1}(13)

Fragment set X1will be assigned to Task 1, as shown

in Fig. 8. The job of the Task 1 model is to calculate

(w1,1, . . . , w1,n). Expanding to arbitrary Task i, the fragment

set Xiwill be offered to Task i to calculate (wi,1, . . . , wi,n ).

The ﬁnal step is to integrate the results of all of the tasks

in order to construct the similarity matrix. An overview of

the method of calculating the similarity matrix in a parallel

manner is shown in Fig. 8.

Step 2: Simplifying the calculation of the eigenvalues of

the graph Laplacians.

Lanczos algorithm [47] is the method used to calculate the

eigenvalues, and the calculation process is shown in Fig. 9.

Based on the process shown in Fig. 9, the following

relationships can be derived:

V0LV = T,V = {v1, v2, . . . , vn}(14)

FIGURE 8: Method of calculating the similarity matrix in a

parallel manner.

FIGURE 9: Method of calculating the eigenvalues of the graph

Laplacians.

T = tridiag(B,A,B),B = {b1, . . . , bn},A = {a1, . . . , an}

(15)

By observing the Lanczos algorithm calculation process,

we ﬁnd that most of calculation time cost is due to the process

of L×vj, so we split L into n rows and multiply each row by

vj. We then merge the results to get the ﬁnal value of L×vj.

An overview of the calculation process is shown in Fig. 10.

The parallel calculation process increases memory cost, but

the inner memory assignment mechanism limits this problem

to a tolerable level.

Step 3: K-means is an iteration process, so we split the data

into several smaller data sets.

We ﬁrst choose random center points for the whole data

set and assign the center points to each data subset. The

subset data will be used to calculate the distance between the

subset data and the randomly chosen center points. Next, the

VOLUME 4, 2016 7

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 10: Parallel calculation of the eigenvalues of the

graph Laplacians.

subset data results are sent to a task which integrates all of

the results of the data subsets, in order to obtain new center

points for the whole data set. This process will continue until

the center points satisfy the demands of the overall data set.

Compared to the traditional K-means process, parallel K-

means calculation converts global calculation into regional

calculation, which simpliﬁes the calculation object in order

to reduce the time cost. The parallel K-means calculation

process is shown in Fig. 11.

FIGURE 11: Method of calculating K-means in a parallel

manner.

FIGURE 12: Driving data clustering results.

3) Feature extraction results

A total of 8,984 natural driving data samples (i.e., the

number of completed trips) were selected during the data

collection process described in Subsection A above. Using

the parallel spectral clustering algorithm described above,

the data samples were then clustered into three groups, with

each group containing drivers with similar driving styles or

behavior, as shown in Fig. 12. The X and Y axes of Fig.

12 represent velocity and positive acceleration, respectively.

The points in the blue cluster represent the drivers who drove

at low velocity with low positive acceleration. The points

in the yellow cluster represent the drivers who preferred to

drive at low velocity but who used high rates of acceleration.

Points in the red cluster represent the drivers who preferred

to drive at a high velocity and whose acceleration ranged

from high to low. We break the clustering results down

statistically using our six selected fuel consumption-related

parameters in Fig. 13. In Fig. 14, the data points of each of

the clusters are plotted on 2-D and 3-D graphs according to

fuel-consumption and their serial number within the data set.

Average fuel consumption for drivers on the blue line was

3.68 L/100 km, on the yellow line 5.14 L/100 km and on the

red line 7.44 L/100 km.

There are several phenomena illustrated in Fig. 13 which

are worth noting. First, we ﬁnd that fuel-consumption within

each cluster differs and that fuel-consumption increases

steadily from the blue to the yellow to the red cluster, i.e.,

there is a surprising amount of variation within each group,

but this variation is constrained by a clear trend. Second,

some outlier points exist, which represent drivers whose fuel

consumption was actually higher than that of some of the

drivers in the next cluster. A numerical analysis of these

outlier points is shown in Table 2. We can see clearly in

the Fig. 14 that the height of each cluster, which represents

increasing fuel consumption, differs. We can also see that the

three clusters have overlapping areas, which can be observed

in the areas of the 3D graph containing blended colors. These

overlapping areas represent the outlier points. Because the

8VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 13: Driving data clustering results for all of the selected parameters (Blue, Yellow and Red refer to the data clusters

shown in Fig. 12).

FIGURE 14: Fuel consumption distribution of the three-cluster group.

TABLE 2: Numerical analysis of outlier points

Group # of outlier

points

Total # of

points

Proportion

of outlier

points

Low fuel

consumption

276 3000 9.20%

Medium fuel

consumption

345 3001 11.49%

spectral clustering process is based on a data graph partition

algorithm, the points on the periphery of each cluster group

will tend towards randomness, which means the points on

the boundaries will join the clusters randomly. Additionally,

the six chosen parameters represent only the major factors

affecting fuel-consumption, but not all of the factors related

to vehicle operation. As a result, some data points which

have high fuel-consumption attributes may also share other

attributes with data points in the lower fuel consumption clus-

ters. What’s more, long-term fuel consumption is deduced by

observing instantaneous fuel consumption, as shown in Table

1, so the calculated fuel consumption values could have an

error rate of 0.8%-5.9%, which could also affect the ﬁnal

clustering results. Finally, the overall proportion of outlier

data points is about 20.69%.

From the above results, we can conclude that drivers who

operate their vehicles with relatively low fuel consumption

are those who change their driving speed moderately and

drive their vehicles at a relatively low average speed. The pro-

posed parallel spectral clustering algorithm was able to accu-

rately cluster the drivers according to their fuel-consumption

using vehicle operation data, with an approximate clustering

accuracy rate of 79.31%.

In order to verify the performance of the clustering method

used in this study, we compared our clustering results with

VOLUME 4, 2016 9

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

those of the kernel fuzzy C-means (KFCM) [30] and K-

means clustering methods [48]. Performance of the three

clustering methods are compared in Table 3.

From this comparison we can see that the proportion of

outlier points when using KFCM is 4% higher than when

using the proposed parallel spectral clustering method. When

using the K-means method, the data points are less tightly

clustered compared with the other two clustering methods,

and the proportion of outlier points is the highest of all the

clustering methods. Therefore, the proposed parallel spectral

clustering method achieved the best clustering performance

of the three methods.

We then compared the calculation efﬁciency of the pro-

posed parallel calculating structure with normal spectral

clustering. Different sample sizes were chosen to verify the

proposed method’s superior performance. The results are

shown in Fig. 15. When the amount of data being calculated

is greater than 10,000 data points, the time cost of normal

spectral clustering using Matlab is almost 18 times higher

than when using the parallel spectral clustering method.

Furthermore, as the amount of data increases, the time cost

of normal spectral clustering increases sharply.

FIGURE 15: Comparison of calculation efﬁciency of classical

and parallel spectral clustering methods.

In this section we described the clustering method used

to obtain the macroscopic relationship between driving be-

havior and fuel consumption. In the next section, an LSTM-

based method is proposed to analyze this relationship in a

more detailed or microscopic way.

III. PREDICTION OF SHORT-TERM FUEL

CONSUMPTION USING LSTM

The clustering-based method proposed in Section II above

can only provide relatively long-term (25 to 35 minutes)

assessment of the impact of a driver’s behavior on fuel

consumption. When attempting to perform relatively short-

term prediction (30 seconds to 5 minutes), the clustering-

based method does not work well for classifying driving

behavior according to fuel efﬁciency. Besides, our clustering

method is, in fact, a kind of classiﬁer, so it has no prediction

ability. Therefore, in this section we propose the use of a

time series learning method (an LSTM network) to model

the relationship between driving behavior and fuel consump-

tion, allowing us to predict the short-term fuel consumption

state of a driver’s behavior. As a driving behavior pattern

represents the driver’s interaction with a dynamic driving

environment, and fuel consumption can be treated as the

cost result of this process, in this section we add dynamic

driving environment information to our learning data. In the

series data construction process described in this section,

we ﬁrst explain how we coded driving environment factors

into a digital form. Then the environmental feature data and

the behavior data are integrated into time-series data using

a sliding window. Fuel consumption state will be the label

for the constructed time-series data set. The LSTM-based

model is then trained using the time-series data. The model’s

classiﬁcation performance and prediction accuracy will be

discussed at the end of this section.

A. TIME-SERIES DATA CONSTRUCTION

1) Coding of environmental factors

As explained in our previous study [49], we divided the

environmental factors into two categories, dynamic envi-

ronmental features (other vehicles, brake lights of leading

vehicles, pedestrians, etc.) and static environmental features

(features which remain invariable for relatively long periods

of time, including road structures such as intersections and

curves). The driving environment factors used for training

our model are shown in Table 4. Some of the dynamic

features are captured by a camera mounted on the vehicle.

As shown in Figs. 2 and 16, two types of roads were used in

this study. In Fig. 16, the gray car is the experimental vehicle,

the red vehicle is the leading vehicle or leading vehicle in

the right lane, the blue vehicle is a parked vehicle, the green

vehicle is the ﬁrst on-coming vehicle in the opposite lane

and the yellow vehicle is the second on-coming vehicle in

the opposite lane. In ordinary-road scenes (one lane in each

direction), the motorcycle or motorbike and the pedestrian

are also considered to be environmental factors which can

affect the driver’s behavior. Thanks to the development of

object detection technology, we can easily extract these traf-

ﬁc environment factors. In this study we used YOLOv3 [50],

a deep learning-based, real-time object detection method, to

obtain the relative positions of these trafﬁc factors. Using this

position information, we can code the trafﬁc factors into a

digital form.

Examples of the raw output of the YOLO network are

shown in the two images on the left of Fig. 17. Environmental

factors beside the road which will not affect driving behavior

are also detected by YOLO. As the camera position is ﬁxed,

a lane detection program can be used to determine lane

position. Using the lane boundary indicator (blue dotted line

10 VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

TABLE 3: Comparison of different clustering methods

Clustering

algorithm

Fuel con-

sumption

group

# of outlier

points

Total # of

points

Outlier point

proportion

Total

proportion

of outlier

points

Spectral cluster Low 276 3000 9.20% 20.69%

Medium 345 3001 11.49%

KFCM Low 417 3117 13.38% 24.53%

Medium 334 2995 11.15%

K-means Low 535 2889 18.52% 28.71%

Medium 327 3210 10.19%

TABLE 4: Driving environment factors considered in the

training data

Dynamic features

On-road traf-

ﬁc factors

Leading

vehicle’s brake

lights

Position of the

vehicle in the

right lane

Position of the

vehicle in the

left lane

Positions of

parked vehicles

Position of

merging vehicle

Positions of

pedestrian &

bicycles

Static features Road structure

Curves

Uncontrolled

intersections

Controlled

intersections

shown in upper left image of Fig. 17), we can remove the

detected environmental factors which are not located within

the range of the road lane. The other noise in YOLO’s output

is the multi-bounding box. We ﬁrst identify the unneeded

multi-bounding boxes by comparing the center points of each

box, and then remove the box with the lower conﬁdence

rating.

After removing the redundant roadside data and the un-

needed bounding boxes, we classify the environmental fac-

tors, using the feature categories listed in Table 4, according

to their positions in the camera image, as shown in Fig. 18.

In our previous study [46], we discovered that providing

the positions of the detected environmental factors helps the

LSTM learn driving behavior more effectively. So, in this

study, we use the same method to change the continuous

positions of trafﬁc objects into discrete locations using a

mapping grid. As shown in Fig. 19, the positions of trafﬁc

factors, such as the vehicles in the photo, are labeled as the

FIGURE 16: Road types and dynamic trafﬁc factors consid-

ered in this study.

FIGURE 17: Correction of raw YOLO output.

VOLUME 4, 2016 11

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 18: The classiﬁed trafﬁc factors label for the object

detected by YOLO.

FIGURE 19: Position zones for identifying the locations of

trafﬁc factors [49].

belonging to an area or zone, in this case areas A2 and B1.

The size of each object is labeled according to the length of

the yellow line under the object.

2) Fuel consumption feature labeling and time series data

construction

In (16), BTrepresents the driving behavior data set from

one trip along the ﬁxed driving route, while S represents

the size of the data (the number of behavior data points)

collected during the time period it took to complete the route.

S is calculated by applying the method shown in Fig. 7

(compression of all of the data sets into the same size). The

only difference in compressing process used in this section is

that here, we divide the experimental road into 150 segments

instead of 50 in order to obtain much more detailed data

features. N in (16) represents the driving behavior categories

strongly and moderately correlated with fuel consumption (N

= 6, which are listed in Fig. 6).

BT=





b1,1··· b1,S

.....

bN,1··· bN,S





N×S

(16)

In (17), ETrepresents the environmental data set from one

trip along the driving route. S is the size (number of envi-

ronmental data points) of the environmental data collected

during the period of time it takes to complete one circuit

of the driving route. M represents the environmental factor

number from the list in Table 3 (M = 13).

ET=





e1,1··· e1,S

.....

eM,1··· eM,S





M×S

(17)

When collecting the driving data, in addition to the camera

frames we also collect the driving behavior data associated

with each frame simultaneously, so that each set of behavior

data corresponds to one camera frame. This allows us to

integrate driving behavior data set BTand environmental

data set ETinto a single dataset XT:

XT=







b1,1··· b1,S

.....

bN,1··· bN,S

e1,1··· e1,S

.....

eM,1··· eM,S





(M+N)×S

(18)

Fuel consumption Fcan then be calculated as follows:

F (XT) = {F1, F2, . . . , Fi, . . . , FI}(19)

Function f(x) in (20) and (21) represents the hypothetical

equation which describes the nonlinear relationship between

driving behavior, driving environment and fuel consumption

features. To deduce function f(x)would be relatively difﬁ-

cult, so here we treat f(x)as a ‘black box’, so our LSTM-

based method is applied to simulate the computations of

this ‘black box’. As the input for the LSTM should be data

in a time-series format, the raw training data must ﬁrst be

converted into time-series data. As shown in Fig. 20, we use

a sliding window to construct each set of time-series data,

and the data label is each data segment’s fuel consumption

Fi. The window size is 50 data points and the size of the

sliding step is 15 data points, so in (21), step = 15 and j = 50.

Fiis mapped into the data segment’s distribution to obtain

its ranking level. For example, in Fig. 20, F1belongs to

the low fuel consumption level (marked with dotted points),

so the “time-series data 1” will be labeled as “low fuel

consumption”. The green, yellow, and red labels represent the

low, medium, and high fuel consumption group respectively.

The fuel consumption group is judged by the other driver’s

12 VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 20: Time series data composition using sliding win-

dow.

historical records.

F1=f







b1,1··· b1,j

.....

bN,1··· bN,j

e1,1··· e1,j

.....

eM,1··· eM,j







(20)

Fi=f







b1,1+(i−1)×step ··· b1,(i−1)×step+j

.....

bN,1+(i−1)×step ··· bN,(i−1)×step+j

e1,1+(i−1)×step ··· e1,(i−1)×step+j

.....

eM,1+(i−1)×step ··· eM,(i−1)×step+j







(21)

The boundaries of the fuel consumption levels are deﬁned

by the trisection lines, and the equation for calculating the

boundaries is shown in (22). Results (0,0,1), (0,1,0) and

(1,0,0) represent low, moderate and high fuel consumption,

respectively. Fiis the current data segment’s fuel consump-

tion and Favg,i is the expected value of the remaining driving

process data in that data segment.

Ii=





(0,0,1), Fi<0.6Favg,i

(0,1,0),0.6Favg,i < Fi<1.2Favg ,i

(1,0,0), Fi>1.2Favg,i

(22)

After completing the labeling process, we can obtain our

training data with fuel consumption feature labels. The labels

are not only obtained by calculating detailed fuel consump-

tion, but also obtained by comparing the fuel consumption

FIGURE 21: The unfolded structure of the LSTM network and

the inner composition of an LSTM node.

distribution with all of the other drivers’ fuel consumption

distributions.

B. FUEL CONSUMPTION PREDICTION MODELING

BASED ON LSTM

1) LSTM components and their mathematical expressions

As the state of the art in information processing and behav-

ior modeling, LSTM is widely used in machine translation

[51], speech recognition [52], driving behavior analysis [53],

and other applications. LSTM is in fact a kind of Recurrent

Neural Network (RNN) [33,54]. Standard RNNs usually suf-

fer from the vanishing gradient problem, but LSTMs include

a ‘forget gate’, which can prevent backpropagation errors

from vanishing or exploding. The structure of the LSTM used

in this study is shown in Fig. 21.

An LSTM is a recurrent network which produces a state

as its output, and the state of current network is passed on to

the next step in the network for further calculation. As shown

in Fig. 21, each node of the LSTM network is composed of

three main components, a ‘forget gate’, an ‘input gate’ and

an ‘output gate’. The ‘forget gate’ determines the effect of

the information from the previous step on the calculations of

the current network, which is the key feature of the LSTM,

allowing it to avoid the problems of gradient vanishing or

exploding. The function of the ‘forget gate’ can be expressed

mathematically as follows:

Fforget =σ(Wf·[yi−1, xi] + bf)(23)

σ(x) = 1

1 + e−x(24)

As σ(x) is a sigmiod fuction, Ff orget is always smaller

VOLUME 4, 2016 13

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

than 1. Furthemore, Fforget will be mutiplied by previous

network state Si−1to form part of the new state Si, so

Fforget determines how much state Si−1will affect current

network state Si.

The second part of the LSTM is the ‘input gate’, which

mainly decides what should be newly added to the current

network state. First, we should ﬁnd which part of the previous

state should be updated, so we use the following equations to

deﬁne the update procedure:

Fin =σ(Wi·[yi−1, xi] + bi)(25)

And then the updated value can be determined as follows:

Snew = tanh (Wnew ·[yi−1, xi] + bnew)(26)

Current network state Sican be obtained from the updated

state value and the remaining previous network state:

Si=Fin ×Snew +Fforget (27)

The third part of the LSTM is the ‘output gate’, which uses

current network state Sito generate the ﬁnal output. Using

current inner state Si, we decide which data we can output,

then the data is multiplied by Fout (which ranges from 0 to

1) to determine which data can be output. The calculation

process is shown in the following equation:

yi= tanh (Si)×σ(Wo·[yi−1, xi] + bo)(28)

In this paper, input xi=XTin (18), and the size of XT,

which is deﬁned by the sliding window in Fig. 20, is 50.

2) LSTM network training process

First, we need to pre-process the training data. All of

the time-series data is normalized into a range of 0 to 1.

We code each data set’s label into a one-hot form: high

fuel consumption is (0,0,1)T, medium fuel consumption is

(0,1,0)Tand low fuel consumption is (1,0,0)T.

To build the LSTM network, we used TensorFlow [55],

which is an end-to-end open source software platform for

machine learning. The LSTM block is based on the LSTM

node unit “tf.nn.rnn_cell.LSTMCell” [56] which is provided

by the TensorFlow API. The hyper-parameters and the train-

ing strategy of the LSTM network are shown in Table 5.

The output of the LSTM is put into a Softmax classiﬁer,

which calculates its probability of belonging to each class.

The Softmax function can convert the output of the LSTM

into a range from 0 to 1. The mathematical expression of the

Softmax function is as follows:

Ci=eyi

Pjeyj(29)

where Ciis the output conﬁdence rate, i.e., the dataset’s

probability of belonging to a certain fuel consumption group.

C. RESULTS OF FUEL CONSUMPTION PREDICTION

USING LSTM

TABLE 5: Hyper-parameters and the training strategy of the

LSTM network.

Hyper-parameters

Parameter

name

Parameter value

Number of

units in the

LSTM’s

hidden layers

125, 150, 200

Number of

hidden layers

in the LSTM

Batch size 64

Initial forget

bias

Initial

learning

rate

0.005

Training strategy Optimizer Adam Optimizer

[57]

Loss function Sparse Softmax

cross entropy

1) Training data

The entire data set is divided into six groups randomly,

with each group containing 5,000 data points of time-series

data. The six groups of data are divided as follows: four

groups are used for training, one group is used for validation

and one group is used for testing. Because the training

process involves cross-validation, each group will be treated

as a training data group, a validation group or a testing group.

2) Comparison of LSTM prediction results with those of

other machine learning methods

We compared the performance of two other machine learn-

ing methods with the performance of the proposed LSTM-

based method. One of those methods was kernel-based Sup-

port Vector Machine (SVM) [58], and the other was a multi-

layer neural network. In addition, LSTM networks with

different number of nodes were also evaluated.

SVM is a very powerful machine learning method which

maps the objects to be sorted into high-dimensional feature

spaces. It is widely used for semantic parsing [59], image

segmentation [60], facial recognition [61] and other applica-

tions. MATLAB’s Statistics and Machine Learning Toolbox

[62] was used to construct our SVM-based classiﬁer.

The multi-layer neural network we used had two hidden

layers, and each layer contained 150 rectiﬁed linear units

(ReLU), as shown in Fig. 22. The output of the network is

passed into a Softmax layer, and the probabilities of the data

belonging to each of the three fuel consumption categories

are calculated.

Two criteria were considered in our evaluation, the clas-

siﬁer accuracy rate and the area under the curve (AUC) of

receiver operating characteristics (ROC) [63]. The classiﬁer

accuracy rate is a direct index which can be used to judge

14 VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 22: Structure of the neural network with two hidden

layers, each of which contains 150 ReLU nodes.

FIGURE 23: AUC values for each modeling method and each

test group.

the performance of the prediction model, however it cannot

evaluate the classiﬁcation performance of the model. AUC

is a probability value, which is the general standard for

evaluating classiﬁer performance. In Fig. 23 we show each

classiﬁer’s performance for each of the six testing groups. We

can see in Fig. 24 that the LSTM with 150 nodes achieved the

best overall performance.

Next, we experimentally evaluated the short-term fuel con-

sumption estimation performance of our proposed LSTM-

based prediction method. Three representative drivers who

belonged to different fuel consumption groups were selected

to test the performance of our deep learning-based predic-

tor. The fuel consumption data distributions for these three

drivers are shown in Fig. 24.

The LSTM-based classiﬁer’s prediction accuracies for

these three drivers are illustrated in Fig. 25.

The red lines represent the predicted fuel consumption

category based on the driver’s fuel consumption features

over time, while the light blue bars represent the actual

distribution of the fuel consumption features corresponding

to the driver’s behavior. The average prediction accuracy for

the three selected drivers was 81%.

FIGURE 24: Fuel consumption data distributions of three

representative drivers.

IV. DISCUSSION AND CONCLUSION

In this paper, we ﬁrst used the unsupervised machine

learning method of spectral clustering to classify drivers into

three groups using six driving behavior-based fuel consump-

tion features. We then analyzed the macro-behavior of each

group, focusing on power demand (speed and acceleration)

and control stability (variation in speed and acceleration).

Our results showed that the proposed spectral clustering-

based method could accurately identify drivers with different

fuel consumption proﬁles, and clearly modeled the relation-

ship between the real-world driving data and the correspond-

ing fuel consumption features.

In addition to the estimation of fuel consumption using ve-

hicle operation data, we also performed a qualitative analyses

of driving behavior, as shown in Fig. 13. Speed and accelera-

tion information reveal the amount of power demanded by a

driver, while variance in speed and acceleration represent the

range of dynamic control exercised by drivers [25, 26]. The

results of our analysis showed that high fuel consumption

drivers (those in the red cluster) tend to maintain a relatively

steady, high demand for power, while their dynamic control

of the vehicle is less stable. Their acceleration rates are higher

and their pedal control behavior is less stable compared to

drivers in the low fuel consumption cluster. Drivers in the

median yellow cluster showed the lowest speed distribution,

but their gas and brake pedal operation characteristics were

similar to those of the low efﬁciency drivers in the red cluster.

Drivers in the blue cluster had the lowest fuel consumption,

since they tended to maintain a consistent speed, and their

dynamic control of the vehicle was the most stable among the

three groups. We also compared the spectral cluster method

VOLUME 4, 2016 15

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 25: Short-term fuel consumption prediction performance using LSTM-based classiﬁer and fuel consumption features

for three representative drivers.

with other state of art clustering method such as k-means

and KFCM. As show in Table.3, spectral cluster method can

achieve the best clustering performance of the three methods.

However, there were drawbacks to our proposed method,

in that the spectral clustering-based method requires rela-

tively long-term data to produce accurate classiﬁcation re-

sults. So, for real-time and short-term fuel consumption fea-

ture prediction, this unsupervised method is not appropriate.

Furthermore, the results of data mining can only show the

impact of a driver’s behavior on fuel consumption on a

macro-level.

Therefore, in the second stage of our study we attempted

to use a supervised machine learning-based LSTM method

to build a link between short-term driving data and the fuel

consumption features. The proposed LSTM-based model was

able to accurately predict driver behavior, achieving a maxi-

mum AUC of 0.836, which is considered to be good human

behavior prediction performance [64]. As shown in Fig.

23, the LSTM-based method achieved better classiﬁcation

performance than the SVM or NN-based methods. LSTM

networks with different numbers of hidden nodes were also

evaluated in this study, revealing that the LSTM with 150

hidden nodes achieved the best average AUC, compared to

LSTMs with 125 or 200 hidden nodes. Three representative

drivers were then selected for a more detailed evaluation of

the model’s performance. As shown in Fig. 25, the short-

term fuel consumption performance of the drivers could be

accurately predicted using the proposed method, although

some prediction error did occur. However, an average overall

prediction accuracy of more than 80% was achieved.The

whole prediction process is end-to-end, as the input of the

model is the driving behavior and dynamic trafﬁc condition

data. After the raw data is reformatted and then processed by

the model, the output is a prediction of which fuel consump-

tion group a particular driver belongs to.

In conclusion, we made three contributions in this paper;

ﬁrstly, we propose a clustering-based data-mining method

which can analyse the behavior and its fuel consumption

result in a macro view. The method can serve as a group

behavior assessment mechanism for the public transporta-

tion department or the commercial transportation company

to evaluate the energy cost distribution. Secondly, we also

propose a micro fuel consumption evaluation model by learn-

ing the driving behavior. The model shows good prediction

ability which can be integrated into the ADAS system or the

eco-driving coach system to evaluate and obtain the fuel-cost

behavior of the single drivers. The predicted state can make

the ADAS or eco-driving system give more reasonable and

adaptive fuel-efﬁcient driving strategy or detail manipulation.

Thirdly, we widen the deep learning method’s application

area, to our knowledge, it is the ﬁrst time that the deep

learning method is used for learning the driving behavior’s

impact on fuel consumption feature.

There are some limitations in our study and in our pro-

posed method. First, the shortcomings of the collected data

will mainly affect the deep-learning based method. As the

collected data are collected from two kinds of road and the

trafﬁc environment factors are not all coded into the time-

series data, so the LSTM can just learn the limited feature

from the ﬁxed trafﬁc condition and the environment it ever

meet. When facing different road types, for example the

road with four lanes, it will suffer prediction performance

decreasing. Second, the prediction accuracy of the proposed

LSTM-based method was not extremely high. We suspect

this is mainly because the model input information included

a limited number of trafﬁc conditions, and because the form

of this input information was relatively basic. As a result, the

LSTM could not accurately predict fuel consumption in very

complex or unknown situations. And our deep-learning based

method is the model can just predict the fuel consumption

level of the driving process so it is hard to give more detail

fuel cost information. What’s more, compared with other

16 VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

state of the art behavior prediction method, LSTM or deep

learning network need lots of training data and training

time. If other new behavior factors which affect the fuel

consumption need to be added into the network, the model

need to revised the original parameter and training process

should be reprocessed. This will limit the generality of the

model. Third, we only used one type of experimental vehicle,

so we need to do further research to determine whether the

proposed LSTM-based model can be adapted to other types

of vehicles. At last, the driver’s personal feature such as age,

sex, driving experiences and so on, are not further studied in

this study.

So, in our future work, ﬁrstly we aim to use larger scale

naturalistic driving data to make our prediction model with

more robustness. Then the other factors’ effect, such as group

personality feature or vehicle type, on the fuel consumption

analysis should also be studied in order to make the fuel

consumption prediction model more general.

ACKNOWLEDGMENT

This work was partially supported by the National Natu-

ral Science Foundation of China [Grant No.61300101] and

key research plan of Jiangsu Province (No.BE2017035).The

authors would like to thank the Vehicle Engineering Devel-

opment Division of Mitsubishi Motors and Functional Safety

Department of UISEE Technologies Beijing Co., Ltd for their

valuable research assistance.

REFERENCES

[1] C. Sun, Y. Luo, and J. Li, “Urban trafﬁc infrastructure investment and air

pollution: Evidence from the 83 cities in China,” J. Clean. Prod., vol. 172,

pp. 488496, 2018.

[2] H. Zhang, S. Wang, J. Hao, X. Wang, S. Wang, F. Chai, and M. Li, “Air

pollution and control action in Beijing,” Journal of Cleaner Production,

vol. 112. pp. 1519-1527, 2016.

[3] World Health Organization, “Health in the green economy. Health co-

beneﬁts of climate change mitigation,” 2011.

[4] H. Li, Y. Yu, X. Qian, D. Fang, Q. Wang, and Y. Lu, “Mortality effects

assessment of ambient PM 2.5 pollution in the 74 leading cities of China,”

Sci. Total Environ., vol. 569-570, pp. 1545-1552, 2016.

[5] H. Liimatainen, “Utilization of fuel consumption data in an eco-driving

incentive system for heavy-duty vehicle drivers,” IEEE Trans. Intell.

Transp. Syst., vol. 12, no. 4, pp. 1087-1095, 2011.

[6] J. E. Meseguer, C. T. Calafate, J. C. Cano, and P. Manzoni, “Assessing

the impact of driving behavior on instantaneous fuel consumption,” in

2015 12th Annual IEEE Consumer Communications and Networking

Conference, CCNC 2015, 2015, pp. 443-448.

[7] C. D’Agostino, A. Saidi, G. Scouarnec, and L. Chen, “Rational truck

driving and its correlated driving features in extra-urban areas,” in IEEE

Intelligent Vehicles Symposium, Proceedings, 2014, pp. 1199-1204.

[8] K. Dietmayer, H. Winner, M. Maurer, K. Bengler, C. Stiller, and B.

Farber, “Three Decades of Driver Assistance Systems: Review and Future

Perspectives,” IEEE Intelligent Transportation Systems Magazine, vol. 6,

no. 4. pp. 6-22, 2014.

[9] J. N. Barkenbus, “Eco-driving: An overlooked climate change initiative,”

Energy Policy, vol. 38, no. 2, pp. 762-769, 2010.

[10] M. J. M. Sullman, L. Dorn, and P. Niemi, “Eco-driving training of

professional bus drivers - Does it work?,” Transp. Res. Part C Emerg.

Technol., 2015.

[11] C. H. Lee and C. H. Wu, “An Incremental Learning Technique for

Detecting Driving Behaviors Using Collected EV Big Data,” Proc. ASE

BigData Soc. 2015, 2015.

[12] J. N`egre and P. Delhomme, “Drivers’ self-perceptions about being an

eco-driver according to their concern for the environment, beliefs on eco-

driving, and driving behavior,” Transp. Res. Part A Policy Pract., 2017.

[13] J. Jim´enez, “Understanding and quantifying motor vehicle emissions with

vehicle speciﬁc power and TILDAS remote sensing,” Massachusetts Inst.

Technol. Cambridge, 1999.

[14] M. Barth, F. An, T. Younglove, G. Scora, C. Levine, M. Ross, and T.

Wenzel, “NCHRP PROJECT 25-11: Development of a Comprehensive

Modal Emissions Model - Final Report,” 2000.

[15] N. Nikkila, M. Osses, J. Lents, N. Davis, and M. Barth, “Development

and Application of an International Vehicle Emissions Model,” in Trans-

portation Research Record: Journal of the Transportation Research Board,

2018, vol. 1939, no. 1, pp. 156-165.

[16] Z. Xu, T. Wei, S. Easa, X. Zhao, and X. Qu, “Modeling Relationship

between Truck Fuel Consumption and Driving Behavior Using Data from

Internet of Vehicles,” Comput. Civ. Infrastruct. Eng., vol. 33, no. 3, pp.

209-219, 2018.

[17] G. Xu, L. Liu, Y. Ou, and Z. Song, “Dynamic Modeling of Driver Control

Strategy of Lane-Change Behavior and Trajectory Planning for Collision

Prediction,” IEEE Trans. Intell. Transp. Syst., vol. 13, no. 3, pp. 1138-

1155, 2012.

[18] Z. Zheng, “Recent developments and research needs in modeling lane

changing,” Transp. Res. Part B Methodol., vol. 60, pp. 16-32, 2014.

[19] H. Xia, K. Boriboonsomsin, and M. Barth, “Dynamic eco-driving for sig-

nalized arterial corridors and its indirect network-wide energy/emissions

beneﬁts,” J. Intell. Transp. Syst. Technol. Planning, Oper., vol. 17, no. 1,

pp. 31-41, 2013.

[20] X. Xiang, K. Zhou, W. Bin Zhang, W. Qin, and Q. Mao, “A Closed-

Loop Speed Advisory Model With Driver’s Behavior Adaptability for Eco-

Driving,” IEEE Trans. Intell. Transp. Syst., vol. 16, no. 6, pp. 3313-3324,

2015.

[21] J. E. Meseguer, C. K. Toh, C. T. Calafate, J. C. Cano, and P. Manzoni,

“Drivingstyles: A mobile platform for driving styles and fuel consumption

characterization,” J. Commun. Networks, 2017.

[22] E. Gilman, A. Keskinarkaus, S. Tamminen, S. Pirttikangas, J. R¨oning, and

J. Riekki, “Personalised assistance for fuel-efﬁcient driving,” Transp. Res.

Part C Emerg. Technol., vol. 58, no. PD, pp. 681-705, 2015.

[23] R. Trigui, S. Javanmardi, E. N. Bourles, H. Tattegrain, E. Bideaux, and

J. F. Tr ´egou¨et, “Driving Style Modelling for Eco-driving Applications,”

IFAC-PapersOnLine, vol. 50, no. 1, pp. 13866-13871, 2017.

[24] C. Lv, X. Hu, A. Sangiovanni-Vincentelli, Y. Li, C. M. Martinez, and

D. Cao, “Driving-Style-Based Codesign Optimization of an Automated

Electric Vehicle: A Cyber-Physical System Approach,” IEEE Trans. Ind.

Electron., vol. 66, no. 4, pp. 2965-2975, 2019.

[25] A. E. af W˚ahlberg, “Long-term effects of training in economical driving:

Fuel consumption, accidents, driver acceleration behavior and technical

feedback,” Int. J. Ind. Ergon., vol. 37, no. 4, pp. 333-343, 2007.

[26] E. Ericsson, “Variability in urban driving patterns,” Transp. Res. Part D

Transp. Environ., vol. 5, no. 5, pp. 337-354, 2000.

[27] M. Ehsani, A. Ahmadi, and D. Fadai, “Modeling of vehicle fuel con-

sumption and carbon dioxide emission in road transport,” Renewable and

Sustainable Energy Reviews, vol. 53. pp. 1638-1648, 2016.

[28] J. Rios-Torres, J. Liu, and A. Khattak, “Fuel consumption for various

driving styles in conventional and hybrid electric vehicles: Integrating

driving cycle predictions with fuel consumption optimization,” Int. J.

Sustain. Transp., 2018.

[29] C. H. Lee and C. H. Wu, “A Novel Big Data Modeling Method for

Improving Driving Range Estimation of EVs,” IEEE Access, 2015.

[30] J. Wu, Y. Du, G. Qi, and M. Xu, “Leveraging longitudinal driving be-

haviour data with data mining techniques for driving style analysis,” IET

Intell. Transp. Syst., vol. 9, no. 8, pp. 792-801, 2015.

[31] Z. Constantinescu, C. Marinoiu, and M. Vladoiu, “Driving style analysis

using data mining techniques,” Int. J. Comput. Commun. Control, vol. 5,

no. 5, pp. 654-663, 2010.

[32] U. Von Luxburg, “A tutorial on spectral clustering,” Stat. Comput., vol. 17,

no. 4, pp. 395-416, 2007.

[33] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural

Comput., vol. 9, no. 8, pp. 1735-1780, 1997.

[34] Q. Han, X. Hu, S. He, L. Zeng, L. Ye, and X. Yuan, “Evaluate Good Bus

Driving Behavior with LSTM,” in Lecture Notes in Computer Science

(including subseries Lecture Notes in Artiﬁcial Intelligence and Lecture

Notes in Bioinformatics), 2018.

[35] S. Kanarachos, J. Mathew, and M. E. Fitzpatrick, “Instantaneous vehicle

fuel consumption estimation using smartphones and recurrent neural net-

works,” Expert Syst. Appl., vol. 120, pp. 436-447, 2019.

[36] B. Degraeuwe, B. Beusen, C. Beckx, T. Denys, L. Govaerts, S. Broekx, M.

Gijsbers, K. Scheepers, L. I. Panis, and R. Torfs, “Using on-board logging

VOLUME 4, 2016 17

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

devices to study the longer-term impact of an eco-driving course,” Transp.

Res. Part D Transp. Environ., vol. 14, no. 7, pp. 514-520, 2009.

[37] SPSS Tutorials: Pearson Correlation. [Online]. Available:

https://libguides.library.kent.edu/SPSS/PearsonCorr

[38] J. Cohen, Statistical power analysis for the behavioral sciences, second

edition. 1988.

[39] P. Ping, W. Qin, Y. Xu, C. Miyajima and T. Kazuya, "Spectral clustering

based approach for evaluating the effect of driving behavior on fuel

economy," 2018 IEEE International Instrumentation and Measurement

Technology Conference (I2MTC), Houston, TX, 2018, pp. 1-6. doi:

10.1109/I2MTC.2018.8409675

[40] M. Stoer and F. Wagner, “A simple min-cut algorithm,” J. ACM, vol. 44,

no. 4, pp. 585-591, 2002.

[41] L. Hagen and A. B. Kahng, “New Spectral Methods for Ratio Cut Parti-

tioning and Clustering,” IEEE Trans. Comput. Des. Integr. Circuits Syst.,

vol. 11, no. 9, pp. 1074-1085, 1992.

[42] J. Shi and J. Malik, “Normalized cuts and image segmentation,” IEEE

Trans. Pattern Anal. Mach. Intell., vol. 22, no. 8, pp. 888-905, 2000.

[43] D. Wagner and F. Wagner, “Between Min Cut and Graph Bisec-

tion”,International Symposium on Mathematical Foundations of Computer

Science, pp. 744-750, 1993.

[44] A. Y. Ng, M. I. Jordan, and Y. Weiss, “On Spectral Clustering: Analysis

and an Algorithm,” in Advances in Neural Information Processing Sys-

tems, 2001.

[45] G. Schoﬁeld, J. R. Chelikowsky, and Y. Saad, “A spectrum slicing method

for the Kohn-Sham problem,” Comput. Phys. Commun., vol. 183, no. 3,

pp. 497-505, 2012.

[46] Apache Spark. [Online]. Available: https://spark.apache.org/

[47] C. Lanczos, “An iteration method for the solution of the eigenvalue

problem of linear differential and integral operators,” J. Res. Natl. Bur.

Stand. (1934)., vol. 45, no. 4, p. 255, 2012.

[48] A. Likas, N. Vlassis, and J. J. Verbeek, “The global k-means clustering

algorithm,” Pattern Recognit., vol. 36, no. 2, pp. 451-461, 2003.

[49] P. Ping, Y. Sheng, W. Qin, C. Miyajima, and K. Takeda, “Modeling Driver

Risk Perception on City Roads Using Deep Learning,” IEEE Access, vol.

6, pp. 68850-68866, 2018.

[50] J. Redmon and A. Farhadi. 2018. “YOLOv3: An incremental improve-

ment.” [Online]. Available: https://arxiv.org/abs/1804.02767.

[51] K. Cho, B.V. Merrienboer, C.Gulcehre, D. Bahdanau, F. Bougares, H.

Schwenk, and Y. Bengio. 2014. “Learning Phrase Representations using

RNN Encoder-Decoder for Statistical Machine Translation.” [Online].

Available: https://arxiv.org/abs/1406.1078.

[52] S. Han, Y. Wang, H. Yang, W. (Bill) J. Dally, J. Kang, H. Mao, Y. Hu, X.

Li, Y. Li, D. Xie, H. Luo, and S. Yao, “Ese: Efﬁcient Speech Recognition

Engine with Sparse LSTM on FPGA” Proc. 2017 ACM/SIGDA Int. Symp.

Field-Programmable Gate Arrays - FPGA ’17, pp. 75-84, 2017.

[53] J. Morton, T. A. Wheeler, and M. J. Kochenderfer, “Analysis of Recurrent

Neural Networks for Probabilistic Modeling of Driver Behavior,” IEEE

Trans. Intell. Transp. Syst., vol. 18, no. 5, pp. 1289-1298, 2017.

[54] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representa-

tions by back-propagating errors,” Cognitive modeling, 1988.

[55] TensorFlow Tutorials, TensorFlow, 2019. [Online]. Available:

https://tensorﬂow.google.cn/tutorials/.

[56] Class LSTMCell,TensorFlow, 2019. [online]. Available:

https://www.tensorﬂow.org/versions/r1.13/api_docs/python/tf/nn/rnn_cell/

LSTMCell?hl=en#class_lstmcell

[57] D. Kingma and J. Ba. 2017. “Adam: A Method for Stochastic Optimiza-

tion” [Online]. Available: https://arxiv.org/abs/1412.6980v8.

[58] C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., vol.

20, no. 3, pp. 273-297, 1995.

[59] R. J. Kate and R. J. Mooney, “Semi-supervised learning for semantic

parsing using support vector machines,” in NAACL-Short ’07 Human

Language Technologies 2007: The Conference of the North American

Chapter of the Association for Computational Linguistics, 2007, pp. 81-

84.

[60] M. Song and D. Civco, “Road Extraction Using SVM and Image Segmen-

tation,” Photogramm. Eng. Remote Sens., vol. 70, no. 12, pp. 1365-1371,

2013.

[61] G. Guo, S. Z. Li, and K. Chan, “Face recognition by support vector ma-

chines,” in Proceedings - 4th IEEE International Conference on Automatic

Face and Gesture Recognition, FG 2000, 2000, pp. 196-201.

[62] Support Vector Machines for Binary Classiﬁcation. [Online]. Available:

https://ww2.mathworks.cn/help/stats/support-vector-machines-for-binary-

classiﬁcation.html.

[63] C. X. Ling, J. Huang, and H. Zhang, “AUC: A better measure than accuracy

in comparing learning algorithms,” in Lecture Notes in Computer Science

(including subseries Lecture Notes in Artiﬁcial Intelligence and Lecture

Notes in Bioinformatics), 2003, vol. 2671, pp. 329-341.

[64] M. E. Rice and G. T. Harris, “Comparing effect sizes in follow-up studies:

ROC area, Cohen’s d, and r,” Law Hum. Behav., vol. 29, no. 5, pp. 615-

620, 2005

PENG PING received the B.S. degree in automa-

tion from Beijing University of Chemical Technol-

ogy, Beijing, China, in 2010 and the M.S. degree

in automation from Nanjing University of Science

and Technology, Nanjing, China, in 2013. From

2013 to 2015, he was a R&D Engineer as part of

the Cloud switch Group in Huawei Technologies

Co.Ltd. He is currently working toward the Ph.D.

degree from Southeast University, Nanjing, China.

From 2017, He went to Nagoya University as a

joint Ph.D. Student. His research interests include vehicle safety, data-

mining, cloud computing and eco-driving.

WENHU QIN received the Ph.D. degree from

Southeast University, Nanjing, China, in 2005.

He is currently a Professor with the School of

Instrument Science and Engineering, Southeast

University, where he has been on its faculty since

1997. He directs the vehicle safety and virtual

reality laboratory at Southeast University. He has

authored or coauthored over 30 journal papers, ten

conference papers, and a book. He is the holder

of three patents. His research interests include

vehicle safety, virtual reality, crowd simulation, and road trafﬁc accident

reconstruction.

YANG XU received the B.S. degree in Instrument

Science from East China University of Technol-

ogy, Nanchang, China, in 2012 and the M.S. de-

gree in Instrument Science from HeFei Univer-

sity of Technology, Hefei, China, in 2015. He is

currently working toward the Ph.D. degree from

Southeast University, Nanjing, China. From 2018,

He went to The University of Queensland as a

joint Ph.D. Student. His research interests include

machine learning, data science, bio-medical signal

processing, and human-computer interaction.

18 VOLUME 4, 2016

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/ACCESS.2019.2920489, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

CHIYOMI MIYAJIMA received the B.E., M.E.,

and Ph.D. degrees in computer science from the

Nagoya Institute of Technology, Japan, in 1996,

1998, and 2001, respectively. From 2001 to 2003,

she was a Research Associate with the Department

of Computer Science, Nagoya Institute of Tech-

nology. She was a Designated Associate Professor

with the Graduate School of Information Science,

Nagoya University, Japan, from 2003 to 2016. She

was an Associate Professor with the Institutes of

Innovation for Future Society, Nagoya University, from 2016 to 2018. Since

2018, she has been an Associate Professor with Daido University, Nagoya,

Japan. Her research interests include the analysis and the modeling of driver

behavior.

KAZUYA TAKEDA received his B.E.E., M.E.E.,

and Ph.D. degrees from Nagoya University, Japan,

in 1983, 1985, and 1994, respectively. Since 1985,

he has worked at Advanced Telecommunication

Research Laboratories and at KDD R&D Lab-

oratories, Japan. In 1995, he started a research

group for signal processing applications at Nagoya

University. His main focus is investigating driving

behavior using data centric approaches, utilizing

signal corpora of real driving behavior. He is cur-

rently a professor at the Institutes of Innovation for Future Society, Nagoya

University. He is also a member of the Board of Governors of the IEEE

Intelligent Transportation Systems Society. He is a Senior Member of the

IEEE.

VOLUME 4, 2016 19

Machine Learning for Real-Time Fuel Consumption Prediction and Driving Profile Classification Based on ECU Data

Article

Full-text available

Jan 2024

Data extracted directly from a vehicle’s electronic control unit (ECU) play a crucial role in the automotive industry because they contain valuable information from the engine and electronic parts. These data have the potential to enable compliance analysis, detect faults and errors, and guarantee driver and car safety as well as product quality. Among the possible uses of the data from the ECUs, driving profile analysis and fuel consumption prediction stand out, which enable analyses for insurers and transportation companies, and help to reduce fuel consumption and greenhouse gases, in addition to providing feedback to the driver. In this work, we apply machine learning algorithms to real data from an engine ECU to examine the driver’s driving behavior and accurately classify their fuel efficiency. Moreover, we develop regression models that predict fuel consumption for vehicles in operation. To ensure the effectiveness of our models, we carefully select variables strongly correlated with fuel consumption using a feature selection process. Compared to related works, both our profile classification results in precision, recall, and accuracy metrics, and our regression models result in the metrics of mean square errors, mean absolute error, and coefficient of determination, which are superior or similar. Notably, our algorithms exhibit lower computational costs and enable real-time analysis by utilizing a cloud server.

Q-learning-based hyper-heuristic framework for estimating the energy consumption of electric buses for public transport

Article

Full-text available

Mar 2024

This research study introduces a Q-learning enhanced hyper-heuristic framework for the accurate estimation of energy consumption rates of electric buses. Fundamentals of reinforcement learning concepts are hybridized with the integrated newly emerged metaheuristic methods of Aquila optimizer, Barnacles Mating Optimizer, Gradient-based Optimizer, Harris Hawks Optimization, and Poor and Rich Optimization algorithms to solve high-dimensional optimization problems with higher accuracy. In this context, the Q-learning algorithm is considered a high-level heuristic for administering the selection and move acceptance mechanisms, while search agents of those mentioned above low-level competitive metaheuristic algorithms meticulously explore the search space to find the optimum global point. Q-learning guides the operating hyper-heuristic in selecting the suitable low-level optimizer based on the Q-table score during iterations. An intelligent control mechanism is devised to get a reward or penalty for the actions of the low-level algorithms. The proposed method is evaluated on thirty-two optimization benchmark problems composed of unimodal and multimodal test functions. Then, each constituent algorithm and the hyper-heuristic model are applied to thirty-dimensional benchmark functions of CEC 2017 and twenty-eight test instances of CEC 2013. Four different challenging, complex real-world engineering design cases are also solved to assess the predictability of the proposed method on constrained problems. Finally, the proposed hyper-heuristic is employed to derive the fuel consumption estimates of electric buses. It is seen that the Multiple linear regression model, whose unknown parameters are extracted by the hyper-heuristic framework, gives the best predictions.

Data-Driven Personalized Energy Consumption Range Estimation for Plug-in Hybrid Electric Vehicles in Urban Traffic

Preprint

Full-text available

May 2024

In urban traffic environments, driver behaviors exhibit considerable diversity in vehicle operation, encompassing a range of acceleration and braking maneuvers as well as adherence to traffic regulations, such as speed limits. It is well-established that these intrinsic driving behaviors significantly influence vehicle energy consumption. Therefore, establishing a quantitative relationship between driver behavior and energy usage is essential for identifying energy-efficient driving practices and optimizing routes within urban traffic. This study introduces a data-driven approach to predict the equivalent fuel consumption of a plug-in hybrid electric vehicle (PHEV) based on an integrated model of driver behavior and vehicle energy consumption. Unlike traditional models that provide point predictions of fuel consumption, this approach uses Conformalized Quantile Regression (CQR) to offer prediction intervals that capture the variability and uncertainty in fuel consumption. These intervals reflect changes in fuel consumption, as well as variations in driver behavior, and vehicle and route conditions. To develop this model, driver-specific data were collected through a driver-in-the-loop simulator, which tested different human drivers responses. The CQR model was then trained and validated using the experimental data from the driver-in-the-loop simulator, augmented by the synthetic data generated from Monte Carlo simulations conducted using a calibrated microscopic driver behavior and vehicle energy model. The CQR model was evaluated by comparing its predictions of equivalent fuel consumption intervals with those of baseline prediction interval methods that rely solely on conformal prediction.

Identification of research areas in fuel sales forecasting within the business ecosystem context: A review, theoretical synthesis, and extension

Article

Full-text available

Apr 2024

Aim/purpose-This paper aims to explore both fuel sales forecasting and the business ecosystem, subsequently reversing the focus to examine the business ecosystem in the context of fuel sales forecasting. Accompanying this research objective are the following research questions: 1) Does the order in which the topics of "business ecosystems" and "fuel sales forecasting" are searched affect the search results? 2) Which keywords frequently co-occur in publications related to "business ecosystems" and "fuel sales forecasting"? 3) What is the relationship between the terms "fuel sales forecasting" and "business ecosystem"? T. Zema, A. Sulich, & M. Hernes 80 Design/methodology/approach-The study employs a hybrid review methodology, utilizing specific queries within the Scopus database to identify research themes and motifs. This hybrid form of literature review integrates the tenets of both bibliometric and structured reviews. In this study, the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework was employed. The visual analysis was conducted using VOSviewer bibliometric software, with a focus on keywords relevant to the relationship between fuel sales forecasting and business ecosystem terms. Findings-Key findings include the identification of co-occurring keywords in fuel sales forecasting and business ecosystem theory literature. The study reveals research gaps and potential areas for future study in business ecosystems, highlighting the impact of fuel sales forecasting in various economic sectors beyond traditional ones, like forestry, agriculture, and fisheries. Utilizing a hybrid literature study research method, the paper analyses data from scientific publications in the Scopus database and employs VOSviewer software to develop bibliometric maps of keyword co-occurrences. Research implications/limitations-The research underscores the broad implications of fuel sales forecasting within a business ecosystem context and identifies areas lacking in-depth study. This study maps scientific publications, identifying the intellectual structure and current research trends. This study contributes to the understanding of fuel sales forecasting within the business ecosystem context as a part of the energy sector transition. Originality/value/contribution-This paper contributes to the field of science and practice by identifying research areas integrating fuel sales forecasting within the business ecosystem construct. It indicates future promising research avenues for researchers and industry professionals, aiming to guide ongoing research. The article addresses a significant theme that warrants scholarly attention. This study allows researchers to define the research gaps covered by published articles and indicate the directions of scientific development.

Application of Data Mining Combined with K-means Clustering Algorithm in Enterprises' Risk Audit

Article

Full-text available

Mar 2024

Dr Sharif Uddin Ahmed Rana

The financial risk management mechanism of enterprises can be more complete through exploration in the application effect of data mining technology combined with K-means clustering algorithm in enterprise risk audit. Hence, K-means clustering algorithm is introduced to study the paperless status of electronic payment in the trading process of e-commerce enterprises. Additionally, a risk audit model of e-commerce enterprises is implemented based on K-means algorithm combined with Random Forest Light Gradient Boosting Machine (RF-LightGBM). In this model, the actual operation process of data preparation, data preprocessing, model construction, model application and evaluation are implemented to study the payment flow in the transaction process of e-commerce enterprises by using big data analysis technology. Eventually, the performance of the model is evaluated by simulation. The results show that, compared with the models and algorithms proposed by scholars in other related fields, the classification accuracy of the model proposed here reaches 95.46 %. Simultaneously, the data message delivery rate of the model algorithm is basically stable at about 81.54 %, and the data message leakage rate, packet loss rate and average delay are lower than those of other models and algorithms. Therefore, under the premise of ensuring the prediction accuracy, the audit model of e-commerce enterprises can also achieve high data transmission security performance, which can provide experimental basis for the safety improvement and risk control of the audit process in e-commerce enterprises.

Driving behavior characterization and traffic emission analysis considering the vehicle trajectory

Article

Full-text available

Jan 2024

Based on the development of the concept of a resource-saving and environmentally friendly society, needing to develop low-carbon and sustainable urban transportation. Most of the pollutants come from the emissions of motor vehicle exhaust. Therefore, this paper analyzes the relationship between driving behavior and traffic emissions, to constrain driver behavior to reduce pollutant emissions. The GPS data are preprocessed by using Navicat for data integration, data screening, data sorting, etc., and then, the speed data are cleaned by using a combination of box-and-line plots and linear interpolation in SPSS. Second, this paper uses principal component analysis (PCA) to downsize 12 indicators such as average speed, average acceleration, and maximum speed and then adopts K-MEANS and K-MEDOIDS methods to cluster the driver’s behavioral indicators, selects the aggregation method based on the clustering indexes optimally, and analyzes the driver’s driving state by using the symbolic approximation aggregation method; finally, according to the above research results and combined with the MOVES traffic emission model to analyze the relationship between the driver’s driving mode, driving state, and traffic emissions, the decision tree can be used to predict the unknown driving mode of the driver to estimate the degree of its emissions.

Recent advances in Machine Learning based Advanced Driver Assistance System applications

Article

Sep 2024
MICROPROCESS MICROSY

Energy consumption analysis of metropolitan logistics vehicles based on an ensemble K-means long short-term memory model

Article

May 2024
Energ Environ

In recent years, creating a green and low-carbon sustainable development has received extensive attention, prompting considerable research into reducing pollution emissions in the transportation sector. This paper analyzes the energy consumption patterns of logistics vehicles on Beijing’s Sixth Ring Road. Firstly, driving segments are categorized based on variations in vehicle speed, followed by the application of the [Formula: see text]-means algorithm for segment clustering, resulting in the identification of three distinct driving states and the construction of corresponding driving cycles. It is observed that the driving states have high correlations with different road grades. Subsequent analysis reveals that speed, torque, and engine speed are the primary factors influencing energy consumption of logistic vehicles. Furthermore, energy consumption prediction models using the long short-term memory algorithm for the identified driving states on various road types are built leveraging historical data, i.e. vehicle speed, motor torque, and engine speed. Finally, the analysis highlights a notable increase in 100 km energy consumption for logistics trucks on branch roads with complex road conditions. This study contributes to the effective management of energy consumption in medium and large trucks.

A Novel Approach Distracted Driver Detection System Using ML Techniques

Article

Jan 2024

Hierarchical Data-driven Predictive Control Strategy for Energy Management of Electric Vehicles

Conference Paper

Dec 2023

Modeling Driver Risk Perception on City Roads Using Deep Learning

Article

Full-text available

Nov 2018

Research on how risk is perceived by drivers is vital to driving behavior research and driving safety. As risk can be divided into subjective and objective risk, in this paper we focus on modeling subjective risk perception by drivers using a deep learning method. Different drivers often perceive different levels of subjective risk under the same driving conditions. In addition, different driving conditions or driving events will have different effects on drivers. Based on these two risk perception features, in this paper we first design an experiment on a city road with two lanes to assess the level of subjective risk perceived by drivers belonging to different groups. We then use a deep learning network based method to abstract features of the driving environment. These environmental features are integrated with driver risk perception data and this information is used as training and testing data for the learning network. Finally, a Long Short Term Memory based method is adopted to model the subjective risk perception of individual drivers based on traffic conditions and vehicle operation data from the driver’s vehicle. Our results show that the proposed method can effectively model the subjective risk perception behavior of drivers, allowing for end-to-end risk perception prediction in future driving assistance systems.

An incremental learning technique for detecting driving behaviors using collected EV big data

Conference Paper

Oct 2015

Instantaneous vehicle fuel consumption estimation using smartphones and Recurrent Neural Networks

Article

Dec 2018
EXPERT SYST APPL

The high level of air pollution in urban areas, caused in no small extent by road transport, requires the implementation of continuous and accurate monitoring techniques if emissions are to be minimized. The primary motivation for this paper is to enable fine spatiotemporal monitoring based on crowd sensing, whereby the instantaneous fuel consumption of a vehicle is estimated using smartphone measurements. To this end, a surrogate method based on indirect monitoring using Recurrent Neural Networks (RNNs) that process a smartphone's GPS position, speed, altitude, acceleration and number of visible satellites is proposed. Extensive field trials were conducted to gather smartphone and fuel consumption data at a wide range of driving conditions. Two different RNN types were explored, and a parametric analysis was performed to define a suitable architecture. Various training methods for tuning the RNN were evaluated based on performance and computational burden. The resulting estimator was compared with others found in the literature, and the results confirm its superior performance. The potential impact of the proposed method is noteworthy as it can facilitate accurate monitoring of in-use vehicle fuel consumption and emissions at large scales by exploiting available smartphone measurements.

Evaluate Good Bus Driving Behavior with LSTM: 5th International Conference, IOV 2018, Paris, France, November 20–22, 2018, Proceedings

Chapter

Nov 2018

Drivers’ behaviors and their decision can affect the probability of the traffic accident, pollutant emissions and the energy efficiency level, good driving behavior can not only reduce fuel consumption, but also improves ride comfort and safety. In this paper, a new concept, evaluation zone, is defined to distinguish special driving areas which has much influence on energy consumption and ride comfort. Then, based on reducing fuel consumption and improving ride comfort, evaluation zone based driving behavior model is proposed to obtain good driving behavior dataset for the long short-term memory (LSTM) to apply the driving behavior evaluation and driving suggestion providing tasks. By using 687# bus line’s driving data of Chongqing City, China, test results demonstrate that the developed model performs well and the LSTM could provide reliable driving evaluations and suggestions for drivers.

Spectral clustering based approach for evaluating the effect of driving behavior on fuel economy

Conference Paper

May 2018

Driving-Style-Based Co-Design Optimization of an Automated Electric Vehicle: A Cyber-Physical System Approach

Article

Jun 2018

This paper studies the co-design optimization approach to determine how to optimally adapt automatic control of an intelligent electric vehicle to driving styles. A cyber-physical system (CPS) based framework is proposed for co-design optimization of the plant and controller parameters for an automated electric vehicle, in view of vehicle's dynamic performance, drivability, and energy along with different driving styles. System description, requirements, constraints, optimization objectives and methodology are investigated. Driving style recognition algorithm is developed using unsupervised machine learning and validated via vehicle experiments. Adaptive control algorithms are designed for three driving styles with different protocol selections. Performance exploration method is presented. Parameter optimizations are implemented based on the defined objective functions. Test results show that an automated vehicle with optimized plant and controller can perform its tasks well under aggressive, moderate, and conservative driving styles, further improving the overall performance. The results validate the feasibility and effectiveness of the proposed CPS-based co-design optimization approach. IEEE

Driving behavior modeling and evaluation for bus enter and leave stop process

Conference Paper

Aug 2017

Fuel consumption for various driving styles in conventional and hybrid electric vehicles: Integrating driving cycle predictions with fuel consumption optimization

Article

Jun 2018

Improving fuel economy and lowering emissions are key societal goals. Standard driving cycles, pre-designed by the US Environmental Protection Agency (EPA), have long been used to estimate vehicle fuel economy in laboratory-controlled conditions. They have also been used to test and tune different energy management strategies for hybrid electric vehicles (HEVs). This paper aims to estimate fuel consumption for a conventional vehicle and a HEV using personalized driving cycles extracted from real-world data to study the effects of different driving styles and vehicle types on fuel consumption when compared to the estimates based on standard driving cycles. To do this, we extracted driving cycles for conventional vehicles and HEVs from a large-scale U.S. survey that contains real-world GPS-based driving records. Next, the driving cycles were assigned to one of three categories: volatile, normal, or calm. Then, the driving cycles were used along with a driver-vehicle simulation that captures driver decisions (vehicle speed during a trip), powertrain, and vehicle dynamics to estimate fuel consumption for conventional vehicles and HEVs with power-split powertrain. To further optimize fuel consumption for HEVs, the Equivalent Consumption Minimization Strategy (ECMS) is applied. The results show that depending on the driving style and the driving scenario, conventional vehicle fuel consumption can vary widely compared with standard EPA driving cycles. Specifically, conventional vehicle fuel consumption was 13% lower in calm urban driving, but almost 34% higher for volatile highway driving compared with standard EPA driving cycles. Interestingly, when a driving cycle is predicted based on the application of case-based reasoning and used to tune the power distribution in a hybrid electric vehicle, its fuel consumption can be reduced by up to 12% in urban driving. Implications and limitations of the findings are discussed.

Development and Application of an International Vehicle Emissions Model

Article

Jan 2005

Rapid vehicle growth in developing nations makes it necessary for these nations to address the transportation and environmental impacts of on-road mobile sources. To estimate the air quality impact of their fleets, many nations have adopted modified versions of U.S. or European emissions models or factors. In most cases, these models can lead to significant errors in emissions estimates. To address this problem, a new on-road mobile source emissions model, called the international vehicle emissions (IVE) model, designed for use in developing countries has been developed. The IVE model was developed jointly by researchers at the International Sustainable Systems Research Center and the University of California at Riverside. The IVE model uses local vehicle technology distributions, power-based driving factors, vehicle soak distributions, and meteorological factors to tailor the model to the local situation. In addition, an intensive 2-week field study was designed to collect the necessary fleet and activity data to populate the model with critical local information. The IVE model, along with the field study process, has proved highly effective in providing an improved estimate of mobile source emissions in an urban area and allows the effective analysis of local policy options. The studies have served to transfer tools and knowledge on the process of creating and improving mobile source inventories in an efficient manner. The rationale behind the development of the model, the development and application of the field studies, an overview of the results obtained to date, and planned next steps are described in this paper.

YOLOv3: An Incremental Improvement

Article

Apr 2018

We present some updates to YOLO! We made a bunch of little design changes to make it better. We also trained this new network that's pretty swell. It's a little bigger than last time but more accurate. It's still fast though, don't worry. At 320x320 YOLOv3 runs in 22 ms at 28.2 mAP, as accurate as SSD but three times faster. When we look at the old .5 IOU mAP detection metric YOLOv3 is quite good. It achieves 57.9 mAP@50 in 51 ms on a Titan X, compared to 57.5 mAP@50 in 198 ms by RetinaNet, similar performance but 3.8x faster. As always, all the code is online at https://pjreddie.com/yolo/

Impact of Driver Behavior on Fuel Consumption: Classification, Evaluation and Prediction Using Machine Learning

Abstract and Figures

Recommended publications

Investigation of User Behavior and Assessment of Typical Operation Mode for Different Types of Firew...

Graph Method Based Clustering Strategy for Femtocell Interference Management and Spectrum Efficiency...

Development of Auto-Modeling Program Based on ADAMS Marcos for Aircraft Landing Gear Dynamic Loads S...

Detection of moving object using a fusion method based on segmentation of optical flow field and edg...