[Paper Review] Traffic Congestion Prediction Using Machine Learning Techniques

Rafed Muhammad Yasir, Asad, Moumita+2 more|arXiv (Cornell University)|Jun 22, 2022

Traffic Prediction and Management Techniques4 citations

TL;DR

This paper proposes a long-term traffic congestion prediction model using Support Vector Regression (SVR) with radial basis function (RBF) kernel, incorporating time, day of week, and weather factors (temperature, humidity, wind speed, etc.) to forecast congestion up to one week in advance. Evaluated on New Delhi traffic data, the model achieved an average RMSE of 1.12, demonstrating feasibility for proactive traffic management despite limitations from a small dataset.

ABSTRACT

The prediction of traffic congestion can serve a crucial role in making future decisions. Although many studies have been conducted regarding congestion, most of these could not cover all the important factors (e.g., weather conditions). We proposed a prediction model for traffic congestion that can predict congestion based on day, time and several weather data (e.g., temperature, humidity). To evaluate our model, it has been tested against the traffic data of New Delhi. With this model, congestion of a road can be predicted one week ahead with an average RMSE of 1.12. Therefore, this model can be used to take preventive measure beforehand.

Motivation & Objective

• To develop a long-term traffic congestion prediction model that accounts for time, day of week, and weather conditions.
• To improve upon existing models that focus only on short-term predictions or omit key factors like weather.
• To evaluate the model’s performance using real-world traffic and weather data from New Delhi.
• To identify the impact of combined temporal and meteorological factors on congestion prediction accuracy.
• To lay the foundation for scalable, data-driven traffic management systems using machine learning.

Proposed method

• The model uses Support Vector Regression (SVR) with a radial basis function (RBF) kernel to map input features into a higher-dimensional space for nonlinear regression.
• Input features include road name, time of day, day of week, temperature, humidity, wind speed, daylight, and speed ratio (current vs. average speed).
• The response variable is the 'jam factor,' a normalized congestion level from 0.0 (free flow) to 10.0 (severe congestion).
• Data was collected at 5-minute intervals from the HERE API for two consecutive weeks (April 15–28, 2019), with the first week used for training and the second for testing.
• Model performance was evaluated using Root Mean Square Error (RMSE), computed as $ RMSE = \left[ \frac{1}{n} \sum_{i=1}^{n} (|f_i - \hat{f}_i|)^2 \right]^{1/2} $.
• The model was compared with AMWR, a short-term prediction method using SVR with adaptive windowing.

Experimental results

Research questions

• RQ1 Can a machine learning model that integrates time, day of week, and weather data predict traffic congestion up to one week in advance with acceptable accuracy?
• RQ2 How do weather factors such as temperature, humidity, and wind speed influence traffic congestion prediction performance?
• RQ3 What is the impact of dataset size on the generalization and accuracy of long-term traffic prediction models?
• RQ4 How does the proposed SVR-based model compare to existing short-term prediction models like AMWR in terms of RMSE and prediction horizon?
• RQ5 To what extent do temporal and meteorological features collectively improve congestion prediction beyond traditional time-based models?

Key findings

• The proposed model achieved an average RMSE of 1.12 across four test locations in New Delhi, indicating moderate predictive accuracy for long-term congestion forecasting.
• The lowest RMSE was recorded at Location 1 (0.893), while the highest was at Location 3 and 4 (1.234 and 1.233, respectively).
• Despite lower RMSE values, the AMWR model outperformed the proposed approach in all locations, achieving RMSEs between 0.521 and 1.939, but only for short-term predictions (up to 15 minutes).
• The model's performance was limited by the small size of the training and testing dataset (one week each), suggesting that data quantity significantly affects prediction accuracy.
• Incorporating weather factors such as temperature, humidity, and wind speed improved model robustness, though not enough to overcome data scarcity.
• The study confirms that combining temporal and meteorological features enhances long-term congestion prediction, even with moderate RMSE, supporting the need for larger datasets in future work.