223978-1-5090-2927-3/16/$31.00 ©2016 IEEE
Performance Evaluation of the Deep Learning
Approach for Traffic Flow Prediction
at Different Times
Yanjie Duan
∗‡
, Yisheng Lv
∗†
and Fei-Yue Wang
∗‡
∗
The State Key Laboratory of Management and Control for Complex Systems
Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China
†
Beijing Engineering Research Center of Intelligent Systems and Technology
Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China
‡
Qingdao Academy of Intelligent Industries
Qingdao, Shandong, 266109, China
Abstract—Traffic flow prediction is very important in the
deployment of intelligent transportation system. Based on our
previous research on deep learning approach for traffic data
prediction, we further evaluates the performance of the SAE
model for traffic flow prediction at daytime and nighttime.
Through 250 experimental tasks training a SAE model and
evaluating its performance at daytime and nighttime with 3
different criteria, we obtain the best combination of hyper
parameters for each criterion at different times on weekday and
non-weekday, respectively. Experimental results show that the
MAE and RMSE at daytime are larger than that at nighttime,
while the MRE at daytime are smaller than that at nighttime.
For different criteria, the hyper parameters of the SAE model
should vary accordingly. The results in this paper indicate that in
real applications, traffic flow prediction using the deep learning
approach can be a combination of multiple SAE models with
different parameters suitable for different periods, which is of
significance in future research.
I. INTRODUCTION
Traffic flow information is very important in the deployment
of intelligent transportation system (ITS) [1]. Accurate and
timely prediction of traffic flow can provide support for road
conditions analysis, dynamic route guidance and traffic signal
control [2]. Thus traffic flow prediction is a critical subject in
ITS research and applications.
There have been many researchers studying traffic flow
prediction using various methods. These methods are mainly
divided into two categories: data-driven methods and model-
driven methods. Data-driven methods rely on traffic flow data
collected from traffic sensors, e.g. cameras, inductive loops
etc. Typically, historical traffic flow data are utilized to build
a prediction model. Then real-time traffic flow data are fed
into the model and the predicted traffic flow in future time is
obtained from the output of the model. Models or algorithms
in these methods include the autoregressive integrated moving-
average (ARIMA) model [3] and its variations [4], [5], Kalman
filtering model [6], k-nearest neighbor (kNN) algorithm [7]
, multivariate regression model [8], neural network model
[9], [10] and deep learning model [11] and so on. Model-
driven methods are based on the principle of dynamic traffic
assignment (DTA). A proper traffic model and the travel
demand data of the road network are essential to conduct
the simulation process of DTA. Through the simulation, the
traffic flow of one specific site can be detected from the virtual
road network. There are many transportation softwares such as
DynaSmart [12], DynaMIT [13], Vissim [14], Paramics [15],
TransWorld [1], which can build virtual road networks for
traffic simulations. However, they all need the travel demand
data named origin-destination (OD) matrix or the demographic
data which must be determined by engineers in advance and
is hard to acquire.
With the growth of traffic data quantity and the improve-
ment of computing capability, data-driven methods [16] have
attracted the attention of many researchers and achieved in-
spiring results. Among them, the deep learning based methods
lead the trend of big data processing. Exploring deep learning
models in the application the traffic flow prediction is of great
significance. We have done the work [17] in which global road
flow are predicted with the deep learning model named stacked
autoencoder (SAE). Huang etc. [18] used the deep learning
model named deep belief network to predict traffic flow. The
existing research demonstrated that deep learning is promising
in traffic flow prediction, and the performance evaluations
focused on traffic flow prediction at all times in a day.
However, we are more concerned about the performance of
the deep learning model at different times in a day. Therefore
in this paper, we further evaluated the performance of the deep
learning model for traffic flow prediction at different times.
Different from our previous research, we discuss the prediction
of traffic flow collected from original vehicle detector stations
(VDSs) rather than the processed global road flow.
The rest of this paper is organized as the following: Section
II describes the SAE model for traffic flow prediction. Section
III presents the experiments conducted on the model using
real traffic flow with different model set, evaluates the perfor-
mances at different times. Section IV concludes this paper.