Forecasting accident frequency of an urban road network: A comparison of four artificial neural network techniques

• Date: 01 November 2018
• DOI: http://doi.org/10.1002/for.2542
• Author: Amir Mohamadian Amiri,Meysam Alizamir,Hamid Behbahani,Reza Imaninasab
• Published date: 01 November 2018

RESEARCH ARTICLE

Forecasting accident frequency of an urban road network: A

comparison of four artificial neural network techniques

Hamid Behbahani

1

| Amir Mohamadian Amiri

2

| Reza Imaninasab

3

| Meysam Alizamir

4

1

Department of Civil Engineering, Iran

University of Science and Technology,

Tehran, Iran

2

Department of Highway Engineering,

Iran University of Science and

Technology, Tehran, Iran

3

Department of Highway Engineering,

Lyles School of Civil Engineering, Purdue

University, West Lafayette, IN, USA

4

Young Researchers and Elite Club,

Islamic Azad University, Hamedan, Iran

Correspondence

Hamid Behbahani, Department of Civil

Engineering, Iran University of Science

and Technology, Narmak, PO Box

1684613114, Tehran, Iran.

Email: behbahani@iust.ac.ir

Abstract

Considerable effort has been made to determine which of the most common

prediction modeling techniques performs best, based on crash‐related data.

Accordingly, the present study aims to evaluate how crashes in the urban road

network are affected by contributing factors. Therefore, in the present paper, a

comparison has been done among four artificial neural network (ANN)

techniques: extreme learning machine (ELM), probabilistic neural network

(PNN), radial basis function (RBF), and multilayer perceptron (MLP). According

to the measures used, including Nash–Sutcliffe (NS), mean absolute error

(MAE), and root mean square error (RMSE), ELM was found to be the most

successful approach in addressing the objectives defined in the present study.

Moreover, not only is ELM the fastest algorithm due to its different structure,

but it has also led to the most accurate prediction. In the end, the RReliefF algo-

rithm was utilized to find the importance of variables used, including V/C, speed,

vehicle kilometer traveled (VKT), roadway width, existence of median, and

allowable/not‐allowable parking. It was proved that VKT is the most influential

variable in accident occurrence, followed by two traffic flow characteristics: V/C

and speed.

KEYWORDS

accident frequency,artificial neural network, extreme learning machine, RReliefF algorithm

1|INTRODUCTION

Traffic accidents are associated with fatality, injuries,

financial losses, and delays as direct costs, and with energy

waste, missed workdays, and economic and psychological

consequences as some of the indirect costs (Nassiri, Najaf,

& Amiri, 2014). With this in mind, many efforts have been

made to lessen the number of accidents on rural and urban

roads.

In recent decades, more researchers have been focusing

on predicting accident frequencies in order to identify and

rank effective parameters in its occurrence. Researchers

have tried to employ more capable approaches and develop

models that more accurately estimate the number of

accidents and better correlate governing variables to acci-

dent frequency.

Generally, accident prediction models can be catego-

rized into two major groups: statistical techniques such

as regression on the one hand and artificial intelligence

(AI) approaches on the other. Single and multivariate

deterministic models, probabilistic models and multiple

logistic models are the most well‐known statistical

approaches. For more than two decades, probabilistic

models such as Poisson and negative binomial (NB) have

drawn most attention in the field of traffic safety. How-

ever, these statistical methods are developed based on

strong assumptions and predefined underlying relation-

ships between variables. Once these assumptions are

Received: 13 September 2017 Revised: 18 June 2018 Accepted: 30 June 2018

DOI: 10.1002/for.2542

Journal of Forecasting. 2018;37:767–780. © 2018 John Wiley & Sons, Ltd.wileyonlinelibrary.com/journal/for 767