Deep spatio-temporal dependent convolutional LSTM network for traffic flow prediction

Area of Science:

• Computer Science
• Artificial Intelligence
• Transportation Engineering

Background:

• Intelligent transportation systems (ITS) and smart cities rely on accurate traffic flow prediction.
• Existing methods struggle with spatial and temporal data imbalances, limiting forecast accuracy.
• Effective traffic prediction aids traffic management, travel planning, and resource allocation.

Purpose of the Study:

• To propose a novel deep learning algorithm, STDConvLSTM, for accurate traffic flow forecasting.
• To address spatial imbalance by introducing a space-dependent attention mechanism.
• To tackle temporal imbalance using a time-dependent attention mechanism.

Main Methods:

• Developed STDConvLSTM, a deep learning algorithm incorporating novel attention mechanisms.
• Implemented a space-dependent attention mechanism to adaptively adjust kernel sizes for spatial features.
• Designed a time-dependent attention mechanism to assign differential importance to historical time steps.

Main Results:

• Achieved good performance on two real-world traffic datasets.
• Demonstrated the effectiveness of the proposed space-dependent attention mechanism in handling spatial imbalances.
• Validated the efficacy of the time-dependent attention mechanism in addressing temporal imbalances.

Conclusions:

• STDConvLSTM effectively overcomes spatial and temporal imbalances in traffic flow prediction.
• The proposed attention mechanisms offer a significant advancement for deep learning in ITS and smart city applications.
• Accurate traffic flow forecasting is crucial for optimizing urban mobility and resource management.