چكيده لاتين
The growing demand for personal transportation has made traffic congestion one of the most significant crises in most major cities worldwide. The ability to predict street traffic volume is immensely helpful in various traffic management and control programs. Accurate traffic volume data plays a key role in important decision-making for urban planning and transportation. Access to this data enables officials to plan infrastructure needs with a clearer vision. However, accurate and cost-effective prediction of traffic conditions remains a challenging problem because vehicle counting is only possible in locations equipped with fixed traffic sensors. In Iran, there is a lack of sufficient infrastructure for collecting up-to-date and precise traffic data across all areas.
Modeling, analysis, and prediction of urban traffic necessitate accurate and comprehensive data. In this research, to compensate for the deficiency or lack of traffic data, routing data from the "Nesh@n" service has been utilized. The Neshan service provides practical web services, offering access to data on distance and travel time between a set of origin and destination points. By performing routing at different times of the day and recording the travel time information, the traffic volume on the edges of the urban road network can be estimated.
In this regard, a set of Machine Learning and Deep Learning methods was used to predict the traffic volume of roads in Region 1 of Isfahan (comprising 251 streets) for future hours. The employed ML algorithms include Random Forest and Extreme Gradient Boosting (XGBoost), and the DL algorithms include Long Short-Term Memory (LSTM) Networks, Bidirectional Recurrent Networks (Bi-RNN), and the Transformer model. Fast Fourier Transform (FFT) was also used to identify the main frequencies of traffic volume changes and for comparison with other methods. Traffic prediction for Region 1 of Isfahan was performed using three approaches: independent street prediction, a single model without considering the graph structure, and a single model with graph structure and spatial dependency. In these models, features such as spatial dependency (sum of travel time of incoming street edges), travel time history, temporal features (hour and day of the week), number of incoming edges, and the number of traffic-influencing city centers (like schools, banks, hotels, municipalities, pharmacies, emergency services, and hospitals) as well as student and total population within a 150-meter radius of each road were used.
Based on the results obtained from the three different modeling approaches, it can be concluded that considering spatial dependencies and the graph structure of the road network plays a decisive role in enhancing traffic volume prediction accuracy. While the first approach with basic features performed well in short-term horizons, the addition of spatial-land use information in the second approach led to a noticeable improvement in the accuracy of the XGBoost and Random Forest models. Finally, leveraging the graph structure and the Bidirectional LSTM model in the third approach, due to its ability to extract complex spatio-temporal patterns, provided the most accurate results with an MAE of 5.77 seconds, an RMSE of 12.69 seconds, an of 0.96, and a MAPE of 8.25%. These findings indicate that combining routing data, spatial information, and graph-based deep learning algorithms can provide an efficient solution for accurate traffic prediction in urban road networks.
