In a paper published in the journal Scientific Reports, researchers investigated the predictability of vehicle travel time and traffic status on a main street in Amman, Jordan.
Many machine learning (ML) algorithms were employed using historical hourly traffic data from Google Maps, including neural networks (NN), gradient boosting (GB), support vector machines (SVM), AdaBoost, and nearest neighbors. Results showed an accuracy of around 98–99% in predicting travel time and traffic status six hours ahead. However, the models differed in identifying the most critical indicators influencing these predictions.
Study: Machine Learning-based Predictive Modeling for Traffic Congestion. Image Credit: metamorworks/Shutterstock
Traffic Congestion Research
Past work in traffic congestion prediction spans various methodologies and focuses on different geographical contexts. Recent studies have explored using artificial neural networks (ANN), Ising models, and vision transformers for congestion prediction on city-wide and street-specific scales. In Jordan, research has delved into predicting traffic volume for rural streets and simulating traffic scenarios in urban areas like Amman. Machine learning (ML) techniques such as linear regression and regression trees have been employed alongside diverse features to forecast congestion levels.
Traffic Prediction Methodology
In the methodology, predictive models are constructed and analyzed for vehicle travel time and traffic status based on historical hourly traffic data. This process enables the prediction of conditions on a given street, specifically Al-Madina Al-Monawara St in Amman, Jordan, six hours ahead.
Historical traffic data was collected from Google Maps for Al-Madina Al-Monawara St, a principal street in Amman, Jordan, covering the period from January 1, 2017, to December 31, 2019. This dataset includes details such as date-time and average vehicle travel time per hour for one direction of the street, spanning from the Suhaib Tunnel to the University Hospital Interchange.
The analysts generated two tables from the obtained data: one for predicting vehicle travel time after six hours and another for predicting traffic status (low, mild, or high) after the same duration. A time series table with 26,252 records and 20 columns was prepared for vehicle travel time prediction, with input features representing travel times for consecutive 19-hour periods and the target feature representing travel time six hours later. A categorical table was constructed for traffic status prediction based on hourly traffic statuses over the same timeframe.
Various ML techniques were employed for predictive modeling, including NN, SVM, nearest neighbors, GB, and adaboost. These methods were chosen based on their efficacy in regression and classification tasks, each having specific advantages in different scenarios. Performance metrics such as mean absolute error (MAE), mean squared error (MSE), accuracy score, precision, recall, F score, and Jaccard similarity coefficient were used to evaluate model performance.
Additionally, feature importance was analyzed using permutation-based methods to identify the most critical predictors influencing the models' predictions. This methodology provides a comprehensive framework for constructing accurate predictive models for vehicle travel time and traffic status, facilitating informed decision-making regarding traffic management and infrastructure planning in urban environments like Amman, Jordan.
Experimental Model Evaluation
The experimental results encompassed constructing and evaluating five predictive models to forecast vehicle travel time and traffic status on a given street in Amman, Jordan, six hours in advance. Employing machine learning algorithms such as AdaBoost, NN, nearest neighbors, SVM, and GB, the models were trained and tested using historical hourly traffic data collected from Google Maps.
Each model's performance was assessed using various metrics, such as MAE, R-squared score, MSE, and feature importance analysis to predict vehicle travel time. The AdaBoost regressor demonstrated exceptional performance with minimal error metrics and stable learning curves, indicating robust predictions. Feature importance analysis revealed critical predictors influencing the models' forecasts, such as previous travel times.
Similarly, the models achieved high accuracy scores for traffic status prediction and exhibited consistent performance across different evaluation metrics. The AdaBoost and NN classifiers accurately classified low, mild, or high traffic statuses. Feature importance analysis highlighted significant predictors, including past traffic statuses, which are crucial for accurate classification.
Overall, the experimental results underscored the efficacy of employing diverse machine learning algorithms to predict vehicle travel time and traffic status, providing valuable insights for traffic management and urban planning in Amman, Jordan. Additionally, the comprehensive evaluation of model performance using various metrics ensured the reliability and robustness of the predictive models.
Conclusion
In summary, the experimental study tested whether vehicle travel time (and traffic status) on a street could efficiently predict six hours based on the previous 19 hours' data. The findings confirmed this hypothesis, influencing how city map applications estimate travel time. While current apps excel in tracking real-time traffic, they offer vague estimates for future trips.
Consequently, the results advocate for integrating predictive models into map applications, enhancing user navigation. Moreover, they encourage traffic authorities to leverage recent traffic data for prompt congestion identification and effective traffic management. Future research aims to extend predictions to the next day for advanced planning and traffic management purposes.
Machine Learning Unveils Satellite Salinity Bias Patterns