IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 1, JANUARY 2022 225
Short-Term Traffic Flow Forecasting Method With
M-B-LSTM Hybrid Network
Qu Zhaowei , Li Haitao, Li Zhihui , and Zhong Tao
Abstract— Deep learning has achieved good performance in
short-term traffic forecasting recently. However, the stochasticity
and distribution imbalance are main characteristics to traffic
flow, and these will bring the uncertainty and induce the
network overfitting problem during deep learning. To deal with
the problems, a new end-to-end hybrid deep learning network
model, named M-B-LSTM, is proposed for short-term traffic flow
forecasting in this paper. In the M-B-LSTM model, an online self-
learning network is constructed as a data mapping layer to learn
and equalize the traffic flow statistic distribution for reducing the
effect of distribution imbalance and overfitting problem during
network learning. Besides, the deep bidirectional long short-term
memory network (DBLSTM) is introduced to reduce the uncer-
tainty problem by forward and reverse contexts approximation
process in the stochasticity reducing layer, and then the long
short-term memory network (LSTM) is used to forecast the next
traffic flow state in the forecasting layer. Furthermore, sufficient
comparative experiments have been conducted and the results
show the proposed model has better ability on solving uncertainty
and overfitting problems than the state-of-art methods.
Index Terms— Short-term traffic forecasting, deep learning,
hybrid network, unbalanced data processing, data stochasticity.
I. INTRODUCTION
T
RAFFIC flow forecasting is an essential component
of intelligent transportation system. Real-time accurate
short-time traffic flow forecasting can give traffic guidance for
traffic participants to select an appropriate travel route, and
help the traffic managers with a reasonable control strategy to
relieve traffic congestion. Traffic flow is a time series data with
strong regularity and cyclicity which are the basis for accurate
forecasting, but the randomness and uncertainty of it increase
forecast difficulty [1]. Thus, short-time traffic flow forecasting
is essential also challenging in transportation management and
research fields.
In the past f ew decades, a lot o f methods have been carried
out to forecast short-term traffic flow, such as auto-regressive
integrated moving average ( ARIMA) [2], Kalman filter [3],
[4], fuzzy theory [5], [6], artificial neural network (ANN) [7].
These methods are proven to be useful in extracting the traffic
flow temporal tendency and predicting the future traffic flow.
Manuscript received September 19, 2018; revised March 16, 2019,
May 29, 2019, October 10, 2019, and March 9, 2020; accepted July 6, 2020.
Date of publication July 29, 2020; date of current version December 28,
2021. This work was supported in part by the National Key Research
and Development Program of China under Grant 2019YFB1600500. The
Associate Editor for this article was M. Zhou. (Corresponding author:
Li Zhihui.)
The authors a re with the School of Transportation, Jilin University,
Changchun 130022, China (e-mail: quzw@jlu.edu.cn; liht16@jlu.end.cn;
lizhih@jlu.edu.cn; 1047706274@qq.com).
Digital Object Identifier 10.1109/TITS.2020.3009725
For the shallow architecture or short-step training strategy
of these methods, the computed forecasting results are still
unsatisfying and need to be more accurate in application. With
the development of deep learning and the increase of traffic
big data, many deep learning models have been carried out
to predict short-term traffic flow. Deep belief network (DBN)
[8], [9], stacked autoencoders (SAE) [10], [11], recurrent
neural networks (RNN) [12], long short-term memory network
(LSTM) [13] and hybrid deep learning network [14], [15], etc.
used multiple-layer architecture to capture deep feature in his-
torical data from multiple levels, a nd achieved more reasonable
forecast performance than superficial pattern learning models.
As usual, the stochasticity and distribution imbalance are
common characteristics to traffic flow. The data stochasticity
will bring the uncertainty to deep network for learning and
forecasting, and the distribution imbalance is the main feature
to traffic flow for that the data statistic distribution is over-
concentration in some frequency bands as shown in Fig. 2.
Data in narrow over-concentration bands will be learned too
many times, which will induce the overfitting problem in learn-
ing stage and decrease the deep learning network discrimina-
tion. In general, oversampling and undersampling are common
solutions to deal with distribution imbalance problem in the
training phase for classification problems [16], and the sam-
pling process is used as an independent preprocessing before
deep network learning, this will induce the deep learning net-
work performance varied with hand-crafted sampling strategy
and network parameters are not overall optimized. Further-
more, the sampling process is an assumption that all traffic
states are sampled in advance for learning and all traffic states
distributions are obtained beforehand. In fact, it is impossible
to get all traffic states and its distributions before learning for
traffic flow’s time-varying and stochasticity. Up to now, the
problems on traffic flow’s stochasticity and distribution imbal-
ance are still confused to the online traffic flow forecasting.
To deal with the problems, a new end to end hybrid deep
learning network model, named M-B-LSTM, is constructed by
integral parameters optimization and self-learning for online
short-term traffic flow forecasting. The model integrates the
data mapping layer, stochasticity reducing layer and forecast
layer together to reduce the ‘hand-crafted’ network implement
effect. In the data mapping layer, an online self-learning proba-
bility distribution mapping network is construc ted to generate
and equalize the traffic flow distribution for improving the
data discrimination and reducing the overfitting problem in
network learning. DBLSTM structure is adopted by forward
and backward contexts approximation process to reduce the
1558-0016 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www .ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: DALIAN MARITIME UNIVERSITY. Downloaded on September 08,2022 at 12:48:55 UTC from IEEE Xplore. Restrictions apply.
剩余10页未读，继续阅读
资源评论
