City-Wide Traffic Flow Forecasting Using a Deep Convolutional Neural Network.

Area of Science:

• Intelligent Transport Systems (ITS)
• Deep Learning Applications
• Urban Planning & Management

Background:

• City-wide traffic flow forecasting is vital for effective traffic management and public safety within Intelligent Transport Systems (ITS).
• Accurate short-term traffic prediction is challenging due to complex factors like road network structure, weather, accidents, and holidays.

Purpose of the Study:

• To propose a novel deep-learning-based multi-branch model, TFFNet (Traffic Flow Forecasting Network), for accurate short-term city-wide traffic flow prediction.
• To effectively model spatial correlations and temporal dependencies in traffic flow data, incorporating external influencing factors.

Main Methods:

• Developed TFFNet, a multi-branch deep learning model utilizing spatiotemporal traffic flow matrices and external factors as input.
• Employed a multi-layer fully convolutional framework to capture hierarchical spatial dependencies and a high-dimensional tensor to extract temporal closeness and periodicity.
• Integrated external factors via a fully connected neural network and fused them with the main model output.

Main Results:

• TFFNet successfully captures simultaneous spatial and temporal dependencies from traffic flow matrices.
• The model demonstrates superior performance compared to traditional traffic flow forecasting methods on the experimental dataset.
• The back-propagation method was used for automatic training to fit complex patterns until convergence.

Conclusions:

• TFFNet offers a robust and effective solution for short-term city-wide traffic flow forecasting.
• The model's ability to integrate spatiotemporal data and external factors enhances prediction accuracy.
• This approach advances the capabilities of Intelligent Transport Systems for better urban traffic management.