Interpretable local flow attention for multi-step traffic flow prediction.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Traffic flow prediction (TFP) is crucial for smart cities.
• Neural network methods show promise but struggle with inflow-outflow relationships, leading to inaccuracy and lack of interpretability.
• Existing models often blend flow data, obscuring specific correlations.

Purpose of the Study:

• To propose an interpretable Local Flow Attention (LFA) mechanism for TFP.
• To develop a novel LFA-ConvLSTM model for capturing complex traffic dynamics.
• To enhance the accuracy, interpretability, and efficiency of traffic flow prediction.

Main Methods:

• Introduced a novel Local Flow Attention (LFA) mechanism that explicitly models correlations between traffic inflows and outflows.
• Developed LFA-ConvLSTM, integrating ConvLSTM, LFA, and a feature aggregation module for comprehensive spatiotemporal feature learning.
• Utilized local attention to reduce computational cost and avoid false attention, unlike global spatial attention methods.

Main Results:

• Achieved significant improvements in prediction performance, reducing RMSE by 3.2%-4.6% and MAPE by 6.2%-6.7% on real-world datasets.
• Demonstrated enhanced interpretability through learned attention weights explaining flow correlations.
• Showcased a 32% increase in prediction speed compared to global self-attention ConvLSTM models.

Conclusions:

• The proposed LFA mechanism and LFA-ConvLSTM model offer a more accurate, interpretable, and efficient approach to traffic flow prediction.
• This method effectively addresses limitations of previous models in handling inflow-outflow relationships.
• The findings contribute to advancing intelligent transportation systems and smart city infrastructure.