Spatio-Temporal Prediction of Freeway Congestion Patterns Using Neural Networks -A Conceptual Approach

Abstract

An accurate prediction of actual traffic conditions on freeways is essential for efficient traffic management, safety, and planning. To this end, the knowledge on which traffic state or more exactly which congestion pattern is prevailing, is the crucial basis for any analysis. In this paper, we propose two models, a standard neural network (NN) and a Long Short-Term Memory (LSTM) neural network, for predicting traffic congestion patterns. We provide a concept containing an overview of the problem statement and its significance, discuss the strengths and weaknesses of the proposed models, and outline the data and methodology that will be used for their development. Likewise, we want to compare the results with a simply explainable mixed logit model. As a data base, we use a rich data set of congestion patterns on the A9 freeway in Germany near Munich. Various explanatory variables are inspected to capture the freeway’s layout, previous congestion patterns, speed, and flow information. Our research underscores the significance of incorporating spatio-temporal patterns when predicting traffic conditions on freeways. This approach leads to a significant enhancement in the accuracy of traffic state prediction.

Keywords: traffic state prediction, neural network (NN), recurrent neural network (RNN), long short-term memory (LSTM), congestion patterns, freeway

Full text

Methodology
Mixed logit model: Prediction if specific congestion pattern occurs or free flowing
traffic for the next time interval
Model MNL: infrastructure effects and information on the existence of
previous congestion patterns [3]
Neural network: Prediction if specific congestion pattern occurs or free flowing
traffic for the next time interval
Model NN: features will be speed, flow, infrastructure effects and
information's to the time of the analyzed day
26 to 30 June 2023 Contact:
Vienna, Austria Barbara Metzger
Spatio-Temporal Prediction of Freeway Congestion Patterns
Using Neural Networks - A Conceptual Approach
Metzger Barbara; Kessler Lisa; Bogenberger Klaus
Technical University Munich
Interpolation of the speed data using the Adaptive Smoothing Method by
Treiber/Helbing, 2002: two traffic-characteristic directions: congestion and free-
flow
Congestion classification method introduced by [2]
•Detection of congestion elements & assignment to one of the congestion
patterns defined by [1]
•Simulation of vehicles using virtual trajectories
•Determination of congestion type depending on speed profile of virtual
trajectories
•Congestion defined below threshold per cell 𝑣𝑐𝑟𝑖𝑡 =40 km/h
For prediction, a list of congestion patterns and their properties is processed:
division of the stretch into space-time cells: 500 m and 1 min including the
information of congestion pattern, local and temporary information
In a Nutshell
•SOTA: Predicting freeway traffic states based on
predicting speeds or traffic volumes
•Congestion on freeways follows patterns:
•Patterns are informative because they propagate in
space-time in different ways
•Data set of congestion patterns on freeway in
Germany
•Development of a neural net (NN) to predict
congestion patterns
•Comparison of results with more explainable mixed
logit model
•More complex NNs are still being worked out
•Biggest challenge: representation of the spatio-
temporal dependency of traffic jams
Congestion
Patterns
•
Jam Wave (short traffic breakdown)
•
Stop and Go
•
Wide Jam (broad congested area)
•
Mega Jam (distinct congestion)
Explanatory
Variables
•
Intersections or ramps
•
Number of lanes
•
Weekday and time
•
Previous traffic conditions (up to 15
minutes)
Sketch of considered road stretch: ramps
(magenta), stationary sensor (dashed green).
Example sketch of a neural network adapted
to our data set
Most likely weekday congestion pattern
(Statistical Method)
Results
Data: Basis, Classification and
Processing
Speed Data from 44 inductive loops with an average
spacing of 1.2 km from a German freeway stretch of
50 km within a period of eight month in 2019
NN
Freeflow
Jam
Wave
Stop & Go
Wide Jam
Mega
Jam
Classification Score
Freeflow
0.994
0
0.004
0.001
0
κ=0.563
mcc= 0.446
Jam Wave
0.341
0.102
0.256
0.260
0.042
Stop
& Go
0.211
0.007
0.734
0.036
0.107
Wide Jam
0.271
0.013
0.084
0.608
0.024
Mega
Jam
0.128
0.005
0.040
0.011
0.816
MNL
Freeflow
Jam
Wave
Stop & Go
Wide Jam
Mega
Jam
Classification Score
Freeflow
0.997
0
0.002
0.001
0
κ=0.884
mcc= 0.890
Jam Wave
0.255
0.743
0.002
0
0
Stop
& Go
0.064
0
0.936
0
0
Wide Jam
0.108
0
0
0.892
0
Mega
Jam
0.033
0
0
0
0.967
(g)
Computation of congestion clusters and
classification of congestion type
Discussion and Conclusion
MNL approach shows good classification rates –see tables above
First NN approaches are applicable and promising to improve the prediction of traffic patterns –Challenge: Represent spatio-temporal
behavior of traffic, especially of congestion.
Acknowledgments
The authors would like to thank Landesbaudirektion
Bayern for providing the data.
References
[1] Karl B., Kessler L., Bogenberger K., Automated Classification of
Different Congestion Types, IEEE ITSC 2019.
[2] Kessler L., Karl B., Bogenberger K., Congestion Hot Spot
Identification using Automated Pattern Recognition, IEEE ITSC
2020.
[3] Metzger B., Loder A., Kessler L., and Bogenberger K., Spatio-
Temporal Prediction of Freeway Congestion Patterns using
Discrete Choice Methods (submitted). EURO Journal on
Transportation and Logistics, 2023.
(a) Interpolated speed distribution (b) Classified congestion clusters
Jam Wave
Stop
and Go
Wide Jam Mega Jam
(a) (b)
Examples of congestion pattern