FASTNN: A Deep Learning Approach for Traffic Flow Prediction Considering Spatiotemporal Features.

Area of Science:

• Intelligent Transportation Systems
• Deep Learning
• Traffic Flow Forecasting

Background:

• Accurate traffic flow forecasting is vital for transportation management and planning.
• Existing deep learning models struggle to independently model spatiotemporal aggregation and feature correlations.
• This limitation can introduce biases, impacting transportation planning decisions.

Purpose of the Study:

• To propose a novel Filter Attention-based Spatiotemporal Neural Network (FASTNN).
• To enhance traffic flow prediction accuracy by addressing limitations in current deep learning approaches.
• To improve the reliability of data used in transportation management and planning.

Main Methods:

• Utilized 3D convolutional neural networks (CNNs) for spatiotemporal dependency extraction.
• Incorporated residual units to prevent network degradation.
• Developed a filter spatial attention module for dynamic spatial weight adjustment.
• Introduced a matrix factorization-based resample module to model feature correlations and reduce redundancy.

Main Results:

• FASTNN demonstrated superior prediction performance on large-scale real-world datasets (TaxiBJ and BikeNYC).
• The model effectively captured spatiotemporal dependencies and feature characteristics.
• Experimental results showed significant improvements over baseline and variant models.

Conclusions:

• The proposed FASTNN effectively addresses the challenges of modeling spatiotemporal aggregation and feature redundancy in traffic flow data.
• FASTNN offers enhanced prediction accuracy, providing a more reliable foundation for intelligent transportation systems.
• The model's ability to dynamically adjust spatial weights and learn feature correlations contributes to its improved performance.