Traffic Congestion Prediction Using Machine Learning Techniques

Moumita Asad, Rafed Muhammad Yasir, Dr. Naushin Nower, Dr. Mohammad Shoyaib

Abstract— The prediction of traffic congestion can serve a congestion [8]. Yisheng et al. proposed a deep-learning-
crucial role in making future decisions. Although many studies based traffic flow prediction method by incorporating spatio-
have been conducted regarding congestion, most of these could temporal relations (dependencies among traffic flows in space
not cover all the important factors (e.g., weather conditions).
We proposed a prediction model for the traffic congestion that and time) [9]. They used a stacked autoencoder (SAE) model
can predict congestion based on day, time and several weather to learn traffic flow patterns. The model was compared
arXiv:2206.10983v1 [cs.LG] 22 Jun 2022

data (e.g., temperature, humidity). To evaluate our model, it with four other models- back propagation neural network,
has been tested against the traffic data of New Delhi. With this random walk forecast method, support vector machine, and
model, congestion of a road can be predicted one week ahead RBF neural network. To evaluate the effectiveness of each
with an average RMSE of 1.12. Therefore, this model can be
used to take preventive measure beforehand. model, three performance indexes were used- MAE, RMSE,
and MRE. Their proposed model achieved an RMSE of 50
I. INTRODUCTION for a 15-minute traffic flow prediction. However, the model
Traffic congestion refers to a condition when travel de- yielded higher errors (e.g., RMSE=138.1) for longer time
mand exceeds the existing road system capacity [1]. It has intervals (e.g., 45 mins).
become a major urban transportation problem [2], [3]. It Lee et al. incorporated weather data (e.g., rainfall, hu-
results in waste of valuable time, more fuel consumption, midity, temperature) to predict traffic congestion [10]. At
emission of pollutant gas, health hazards (lung diseases, high first, they developed a multiple linear regression (MLR)
blood pressure) and low productivity at the workplace. World model using 54 variables. Next, they filtered-out unimportant
Bank’s study revealed that around 8 billion USD are wasted variables. The final model consists of 10 variables among
every year in the Greater Cairo Metropolitan Area(GCMA) which, 6 variables represent the days of the week and 4
due to traffic congestion [4]. Drivers in Los Angeles spent are weather factors. Their approach yielded an accuracy of
more than 100 hours in traffic [5]. City dwellers in rich 75.5%. However, they did not consider the time of the day,
countries (e.g., New York, Los Angeles) lose approximately which is an important factor for predicting traffic congestion
$1,000 a year while sitting in traffic [5]. Therefore, a traffic [7].
prediction model is needed to take preventive measures to Akbar et al. combined Complex Event Processing (CEP)
avoid it [6]. with Machine Learning (ML) to predict traffic congestion
Predicting traffic congestion is a challenging task since it [6]. They proposed an algorithm named Adaptive Moving
shows nonlinear, time-varying characteristics [7]. In addition, Window Regression (AMWR) which uses SVR with radial
many uncertain factors such as weather, time and day have based kernel function to predict intensity (number of vehicles
an impact on traffic congestion, which makes it difficult to per hour) and speed based on real-time and historical data.
predict. Therefore, a model is required that incorporates all It uses Lomb Scargle method to find the optimum training
of these factors for predicting traffic congestion. Although window size. For ensuring a certain level of accuracy (80%-
a number of traffic prediction models have been proposed, 95%), AMWR adjusts the prediction window size. If the ac-
most of the models could not cover all these factors. Further- curacy of the model is high (95%>), the prediction window
more, some of the models are only suitable for short-term size is increased. On the other hand, the prediction window
prediction (e.g., 15 minutes). In this paper, we proposed a size is decreased when the accuracy of the model is low
model by combining time, day and weather factors (e.g., (<80%). Lastly, the proposed technique uses CEP rules to
temperature, rainfall) for long-term (one week ahead) traffic predict traffic congestion based on the predicted intensity and
prediction. We validated the model using traffic data of New speed. Although this technique achieved 96% accuracy, the
Delhi, India. The proposed model yielded an average RMSE model mainly focuses on short-term traffic prediction.
of 1.12.
III. M ETHODOLOGY
The rest of the paper is organized into different sections
as follows. Section II presents on literature review. Section In this paper, an approach was developed to predict traffic
III discusses the methodology. Evaluation of the approach is congestion one week ahead, based on the traffic and weather
discussed in section IV. Section V presents the concluding data of the previous week. For this purpose, the traffic and
remarks. weather data of New Delhi, India, collected from the HERE
API were used [11]. To train the model, Support Vector
II. LITERATURE REVIEW Regressor (SVR) algorithm was used since SVR achieved
Due to the increasing necessity for traffic predictive tools, superior performance in traffic prediction than other methods
various approaches have been proposed to predict traffic [7], [12].
A. Data Collection The predicted traffic congestion and the actual traffic
To predict traffic congestion, traffic data of New Delhi congestion of the four roads are shown in Fig.1. Since we
was collected using the HERE API [11]. The bounding box incorporated important factors (e.g.,weather, time and day)
coordinates of the area, whose traffic data were collected, are that have an impact on traffic congestion, it was expected that
(28.747193,77.091064) and (28.495247,77.304611). HERE our model would yield low RMSE. However, the proposed
also provides a weather api [13], through which weather approach could not achieve satisfactory result due to small
data of the selected region was collected. Following existing dataset (1-week training data and 1-week testing data).
literature [6], [8], traffic and weather data was collected at
an interval of 5 minutes. Table I summarizes the features
present in the dataset.
TABLE I: Predictor and Response Classes, Names and
Description
Type Name Description
Source Road from where the traffic is mea-
Predictor sured
Destination Road upto where traffic is measured
Time of a day Time of the day when the data was
collected
Day of the week Represents weekday
Temperature Temperature in degree celsius (a) Location 1 (RMSE=0.893)
Daylight Whether daylight existed or not
Humidity Humidity in percentage
Wind speed Wind Speed in Km/h
Speed ratio Ratio of current speed and average
speed without traffic
Jam factor A number between 0.0 and 10.0 indi-
Response cating the traffic level. As the number
approaches 10.0 the quality of travel
is getting worse.

B. Development of Prediction Model

In this paper, an approach was developed to predict traffic
congestion one week ahead, based on the traffic and weather
data of the previous week. Several algorithms (e.g., LSTM,
(b) Location 2 (RMSE=1.120)
Random Forest Regressor, Gradient Boosting Regressor [14])
can be used for time series regression. We used support
vector regression (SVR) since SVR achieved superior perfor-
mance in traffic prediction than other methods [7], [12]. SVR
is a supervised learning algorithm based on the computation
of a linear regression function in a high dimensional feature
[15]. The input data is mapped to a higher dimension via a
nonlinear function called kernel [16]. Following an existing
approach [6], we used the radial basis function kernel (rbf)
for data transformation.
IV. EXPERIMENTAL EVALUATION
For evaluating our model, we collected traffic data of two (c) Location 3 (RMSE=1.234)
consecutive weeks at an interval 5 minutes each day. The
data of the first week (15.4.2019 to 21.4.2019) was used as
training data and the second week (22.4.2019 to 28.4.2019)
was used as testing data. To evaluate the performance of
the proposed approach, four roads were selected randomly
to measure Root Mean Square Error (RMSE), based on
equation (1).
n 
h1 X 2 i 12
RM SE = |fi − fˆi | (1)
n i=1
where,
(d) Location 4 (RMSE=1.233)
fi = the actual traffic congestion
f̂i = the predicted traffic congestion Fig. 1: Prediction result on different roads
 To compare performance of the proposed approach with [4] “World bank. 2013. cairo traffic congestion study :
existing approaches, we implemented AMWR, which also Final report. washington, dc. © world bank.” https://
openknowledge.worldbank.org/handle/10986/18735.
uses SVR as its underlying method [6]. The results are shown Accessed on: 2018-12-25.
in Fig.2 and Table II. Although AMWR achieved better result [5] “The hidden cost of congestion.” https://www.economist.com/
than our approach in most of the cases, it can predict short- graphic-detail/2018/02/28/the-hidden-cost-of-
congestion. Accessed on: 2018-10-26.
term traffic only (utmost next 15 minutes). On the other hand, [6] A. Akbar, A. Khan, F. Carrez, and K. Moessner, “Predictive analytics
the proposed technique can forecast traffic upto one week for complex iot data streams,” IEEE Internet of Things Journal, vol. 4,
ahead. no. 5, pp. 1571–1582, 2017.
[7] M. Deshpande and P. Bajaj, “Performance improvement of traffic flow
prediction model using combination of support vector machine and
rough set,” International Journal of Computer Applications, vol. 163,
no. 2, pp. 31–35, 2017.
[8] W. Min, L. Wynter, and Y. Amemiya, “Road traffic prediction with
spatio-temporal correlations,” in Proceedings of the Sixth Triennial
Symposium on Transportation Analysis, Phuket Island, Thailand (June
2007), vol. 65, p. 85, 2007.
[9] Y. Lv, Y. Duan, W. Kang, Z. Li, F.-Y. Wang, et al., “Traffic flow
prediction with big data: A deep learning approach.,” IEEE Trans.
Intelligent Transportation Systems, vol. 16, no. 2, pp. 865–873, 2015.
[10] J. Lee, B. Hong, K. Lee, and Y.-J. Jang, “A prediction model of
traffic congestion using weather data,” in 2015 IEEE International
Conference on Data Science and Data Intensive Systems, pp. 81–88,
IEEE, 2015.
[11] “What is the traffic api?.” https://developer.here.com/
documentation/traffic/topics/what-is.html. Ac-
cessed on: 2018-10-27.
[12] J. Ahn, E. Ko, and E. Y. Kim, “Highway traffic flow prediction using
support vector regression and bayesian classifier,” in 2016 Interna-
tional Conference on Big Data and Smart Computing (BigComp),
pp. 239–244, IEEE, 2016.
Fig. 2: Comparison between Proposed Approach and AMWR [13] “Overview - destination weather api.” https://
developer.here.com/documentation/weather/topics/
overview.html. Accessed on: 2018-10-27.
[14] “How to not use machine learning for time series forecasting.”
TABLE II: Comparison between Proposed Approach and https://towardsdatascience.com/how-not-to-use-
machine-learning-for-time-series-forecasting-
AMWR avoiding-the-pitfalls-19f9d7adf424. Accessed on:
Location RMSE of Proposed Approach RMSE of AMWR 2019-05-09.
Location 1 0.893 0.521 [15] D. Basak, S. Pal, and D. C. Patranabis, “Support vector regression,”
Location 2 1.120 0.541 Neural Information Processing-Letters and Reviews, vol. 11, no. 10,
Location 3 1.234 0.735 pp. 203–224, 2007.
[16] “Support vector regression.” https://medium.com/
Location 4 1.233 1.939
coinmonks/support-vector-regression-or-svr-
8eb3acf6d0ff. Accessed on: 2019-05-09.

V. CONCLUSION
In this paper, a traffic prediction model based on sup-
port vector regression with a radial basis kernel has been
proposed. Apart from using features like road name, time
and weekday, weather factors were also used for training the
model as they are important factors that contribute to traffic
congestion. The model was evaluated on four randomly
selected road. It yielded an average RMSE of 1.12.
Due to small dataset, the model could not achieve expected
accuracy. In future, accuracy of the model will be improved
by increasing the size of the dataset. In addition, real time
and historical data will be combined to further improve the
accuracy.

R EFERENCES
[1] S. Rosenbloom, “Peak-period traffic congestion: A state-of-the-art
analysis and evaluation of effective solutions,” Transportation, vol. 7,
no. 2, pp. 167–191, 1978.
[2] A. Downs, Stuck in traffic: Coping with peak-hour traffic congestion.
Brookings Institution Press, 2000.
[3] T. Litman, “Transportation cost and benefit analysis,” Victoria Trans-
port Policy Institute, vol. 31, 2009.