Driving Strategy of Heavy Haul Train based on Support Vector

Regression

Shuo Yang

1, a

, Xiaofeng Yang

1, b

and Zhengnan Lin

1, c

School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100000, China.

Keywords: Heavy Haul Train; Driving Strategy; Support Vector Regression (SVR).

Abstract: In order to reduce the labor intensity of heavy-haul train drivers in the downgrade section and near the split

phase area, the paper analyzes the factors that affect the driving strategy of heavy-haul train in accordance

with the manual driving strategy. In this paper, A control model of heavy-haul train electric braking force

based on support vector regression (SVR) is proposed to control the electric braking force. With the manual

driving records used as training data, Electric braking force and other information are extracted as output

results and features to train the control model. By trial and error, parameters of the control model are adjusted

to optimize the model. The results show that the control model in this paper is close to the manual driving in

the same situation, which is positive for reducing the labor intensity of drivers in heavy-haul railway.

1 INTRODUCTION

In freight transportation, the heavy-haul railway has

the advantages of large capacity, high efficiency and

low transportation cost, which is of great significance

to the "west to East Coal Transportation" project in

China.

Due to the difficulty of heavy-haul train driving,

the drivers need to give their whole attention to

driving with no distractions for a long time; on the

other hand, the heavy-haul line is long, the drivers

need to drive without mistake for more than eight

hours, which cause the high labor intensity of drivers.

In the relevant research of heavy-haul trains,

Wang et al. (Xi Wang, et.al, 2018) had studied the air

braking of heavy-haul trains on the long and steep

downgrade to ensure the safety of heavy-haul trains

in the implementation of air braking; Yu et al. (H.Yu,

et.al, 2018) Put forward an intelligent optimization

method based on particle swarm optimization (PSO)

to generate driving strategy; Lin et al. (Xuan Lin, et.al,

2019) Analyzed the operation energy consumption

through the maximum value principle, gave the

method of energy saving through finding the time of

"full air breaking"; Gao K et al. (Gao K, et.al, 2013)

Designed a distributed controller to solve the control

problem of multiple locomotives in the complex

terrain and unreliable communication of the heavy-

haul combined train.

In addition, some scholars control the train

operation by improving PID algorithm, and Chang et

al. (Chang C, et.al, 2017) Designed ATO controller

by fuzzy differential evolution algorithm to optimize

train operation. Hou et al. (Hou Zhongsheng, et.al,

2011) Used model-free adaptive control method to

stop the train automatically when entering the station.

Shao (Shao, H, 2016) studied the method based on

genetic algorithm (GA). This method has strong

robustness, the dynamic and stability characteristics

of the system have been greatly improved, and the

PID parameters will change with the external

interference.

Huang et al. (Huang Y, et.al, 2016) Designed a

Back Propagation (BP) neural network to generate

driving curve for heavy-haul trains based on genetic

algorithm (GA). The reliability and feasibility of the

method were verified by comparing with the actual

driving curve.

Lu X. et al. (Lu Xiaohong, et.al, 2017) used fuzzy

control to track the recommended speed of heavy-

haul train and obtained a satisfactory result.

Bonissone et al. (Bonissone P P, et.al, 1996)

generated a fuzzy controller to track the speed curve,

and used genetic algorithm (GA) to optimize the

performance of the fuzzy controller by adjusting the

parameters of the fuzzy controller.

Qin et al. (Yufu Qin, et.al, 2014) Designed an

error detection estimator for generating error

detection residuals, and designed the driving strategy

Yang, S., Yang, X. and Lin, Z.

Driving Strategy of Heavy Haul Train based on Support Vector Regression.

DOI: 10.5220/0010121200700074

In Proceedings of the International Symposium on Frontiers of Intelligent Transport System (FITS 2020), pages 70-74

ISBN: 978-989-758-465-7

of the train considering the error mode, which was

verified in Da Qin railway.

This paper mainly focuses on the driving strategy

of heavy-haul train near the phase separation area and

in the downgrade section. First of all, by analyzing

the influence of various line conditions on the driving

strategy of heavy-duty train, the features that affect

the driving strategy are constructed. Then, the

training model is constructed by using the SVR

method with the features and the electric braking

force which is in the manual records. and the model

is optimized by adjusting the parameters of SVR.

2 DATA PREPROCESSING

2.1 Feature Analyze

The information recorded by on-board recording

system can be divided into train formation

information and train status information.

Train formation information includes: total train

weight, number of vehicles, train length, effective

load and so on. These data record the information

related to the train body.

The train status information includes: speed,

kilometer mark, front signal status, distance from

front signal, time, pipe pressure, handle position,

traction force, electric braking force. These data

record the status of the train and operation of the

driver.

In the recorded data, the size of electric braking

force is taken as the result, and other data are included

in the training features.

2.2 Feature Supplement

For the purpose of more accurate prediction of the

driving strategy of the train, the line features near the

running position of the characteristic train are

supplemented from the static data of the line.

When the heavy-haul train is running on the

downgrade, the slope in front of the train is very

important. The slope value 200m, 400m, 600m,…,

1600m in front of the train at 1600m is included in the

training features.

The electric braking force of the train is related to

the distance to the phase separation area, so the

distance between the train and the nearest phase

separation area in the driving process (passing

through is positive, not passing through is negative)

is also added as a set of training features.

Other information such as curvature of the current

position curve of the train, direction of the current

position curve of the train and current speed limit are

also used as training features.

3 ALGORITHM FLOW

3.1 Feature Analyze

When loss function of support vector machine (SVM)

is changed to make SVM used in regression analysis,

it is called support vector regression (SVR).

The problem of SVR can be expressed as follows:

for given training data

[

]

(, ), ,( , )

xy x y= 

, we

hope to get a regression model

, to make the

predicted value

()

as close as possible to the

actual value

Because the trained samples are not necessarily

linearly separable, they can be mapped to high-

dimensional feature space by nonlinear mapping

()



xx

. In this case, the model corresponding to

the hyperplane divided in the feature space can be

expressed as:

() ()





xwx

(1)

In formula (1),

is the normal vector of the

hyperplane, and

is the displacement term, which

determines the distance between the hyperplane and

the origin.

Assuming that the most tolerable deviation is



the problem can be written as follows:

min ( ( ) )

Clf y



(2)

Where

is the penalty factor and



is the

insensitive loss function:

()



-³

(3)

The solution of the original problem can be

expressed as follows:

 

() ( ), ( )

ii i i

 







xxx

(4)

Driving Strategy of Heavy Haul Train based on Support Vector Regression

The calculation method of parameter

is:

 

(),()

iiiii





 



(5)





(),()

 

is the kernel function. The

parameter can be got by calculating the average

value of the trained samples satisfying condition





3.2 Feature Analyze

In order to accurately evaluate the deviation degree of

electric braking force error which is used in this

paper. The evaluation indexes of commonly used

prediction models are as follows:

1. Mean Average Error (MAE)

AE y y







2. Root Mean Square Error (RMSE)

()

RMSE y y







Among the above evaluation indexes,

is the

number of test data;

and

are the actual and

predicted values of the group

of test data

respectively.

3.3 Model Training and Optimizing

Algorithm: Training model and optimizing

1. Use the original data

and the line data

to construct the feature

, according to the

method in Chapter 2.

2. Extract the control data

of heavy-haul train

from the original data

. The extracted dataset





3. The dataset





is divided into





train train

and





test test

according

to the ratio of 4:1. Specify the parameter penalty

factor C and kernel function





(),()

 

SVR, and then train





train train

. Suppose

the training result is

()

train train

YfX

()

test test

YfX

is calculated by using the

control model, and then MAE and RMSE of the

model are calculated. Repeat step 3 and select the

one with the minimum MAE and RMSE as the

optimal control model

.The result

is the

control model.

5. Select some testing data, and compare the

predicted value

test

with the real value

test

4 ALGORITHM FLOW

Firstly, according to the data processing method in

Chapter 2, 10426 data of Shuo Huang railway

operation records are selected as training original

data. The method in Chapter 3 is used for

optimization. Bring equation (6) into the kernel

function of Chapter 3, and optimize the training

model by adjusting γ of equation (6) and penalty

factor C of Chapter 3.

(, )

xx e









(6)

The debugging range of γ in equation (6) and

penalty factor C in Chapter 3 is shown in .

Table 1: SVR debugging scope.

Debugging value Debugging scope

Coefficient of kernel function γ [10-4,10-1]

penalty factor C 500, 800, 1000

FITS 2020 - International Symposium on Frontiers of Intelligent Transport System

MAE and RMSE are as follows:

Figure 1: Optimization of the MAE.

Figure 2: Optimization of the RMSE.

From Figure 1 and Figure 2, when C = 1000 and

γ = 0.0138, the absolute mean error (MAE) and root

mean square error (RMSE) can be minimized. At this

time, MAE = 14.1 (KN), RMSE = 37.2 (KN).

Therefore, C = 1000, γ = 0.0138 is taken as the result

of parameter optimization.

After parameter optimization, some manual

driving records running near the kilometer mark 140-

185 are selected in order to reflect the prediction

results intuitively. The output braking force of the

control model with the selected parameters is

compared with the actual value. The train load 5000

ton with 66 vehicle pulled by SS4B electric

locomotive.

The comparison of predicted and actual values is

shown in Figure 3. It can be seen from the figure that

in the same railway line situations, the driving

strategy based on SVR proposed in this paper is close

to the manual driving operation.

Figure 3: Comparison of predicted the actual values.

5 CONCLUSIONS

In this paper, the driving strategy of heavy-haul train

in the downgrade section and near the split phase area

is studied. The control model of electric braking force

is built by training the manual driving records with

SVR method. After parameter optimization, the MAE

and RMSE of the predicted value and the actual value

output are 14.1kN and 37.2kN respectively.

Therefore, the control model based on SVR in this

paper can be used to reduce the labor intensity of the

heavy-haul train drivers with acceptable error in the

downgrade section and near the split phase area.

REFERENCES

Bonissone P P, Khedkar P S, Chen Y. Genetic algorithms

for automated tuning of fuzzy controllers: a

transportation application. Proceedings of IEEE 5th

International Fuzzy Systems. IEEE. New Orleans, LA,

1996.9.11, p. 674-680.

Chang C, Guo G, Wang J, et al. Study on longitudinal force

simulation of heavy-haul train. Vehicle System

Dynamics. Vol. 55(2017) No.4, p. 571-582.

Gao K, Huang Z W, Wang J, et al. Decentralized control of

heavy-haul trains with input constraints and

communication delays. Control Engineering Practice.

Vol. 21(2013), p. 420-427.

Hou Zhongsheng, Yi Wang, Chengkun Yin, et al. Terminal

iterative learning control based station stop control of a

train. International Journal of Control. Vol. 84(2011)

No. 7, p. 1263-1274.

Huang Y, Tan L, Chen L, et al. A neural network driving

curve generation method for the heavy-haul train.

Advances in Mechanical Engineering. Vol. 8(2016) No.

Driving Strategy of Heavy Haul Train based on Support Vector Regression

H.Yu,Y.Huang, M.Wang. Research on operating strategy

based on particle swarmoptimization for heavy haul

train on long down slope. 2018 21st International

Conference on Intelligent Transport at ion Systems

(ITSC). Maui, HI. 2018.11.4, p. 2735-2740.

Lu Xiaohong, Zheng Muhuo, Lin Hongquan, et al.

Research on and Implementation of Intelligent Control

Algorithm for Heavy Haul Train. Journal of the China

Railway Society. Vol. 39(2017) No. 1 p. 11-18.

Shao, H. Research on PID control parameter tuning based

on genetic algorithm. Wireless Internet Technology.

(2016) No.21 p. 111-112.

Xi Wang, Shukai Li, Tao Tang, et al. Intelligent operation

of heavy haul train with data imbalance: A machine

learning method. Knowledge-Based Systems. Vol.

163(2018), p. 36-50.

Xuan Lin, Qingyuan Wang, Qiangqiang Liu, et al. On

Periodic Braking of Freight Trains on Long Steep

Downhill Section. Journal of the China Railway

Society. Vol. 41(2019) No. 01, p. 50-58.

Yufu Qin, Jun Peng, Xiaoyong Zhang. A robust fault

estimation scheme for Heavy-haul trains equipped with

ECP brake systems. 26th Chinese Control and Decision

Conference (CCDC). Changsha, China, 2014.5.31, p.

2831-2836.

FITS 2020 - International Symposium on Frontiers of Intelligent Transport System