A hybrid deep learning model combining multi-scale self-attention and LSTM for accurate traffic flow prediction

This study introduces a novel traffic flow prediction model integrating Multi-Scale Self-Attention (MSA) with Long Short-Term Memory (LSTM) networks to address the challenges faced by traditional deep learning approaches in managing complex traffic data. The suggested model improves the accuracy of predictions by using MSA to move attention between different time scales in a way that captures both steady patterns and sudden changes. The LSTM component ensures the learning of long-term dependencies, fitting the sequential nature of traffic data.The model was trained and evaluated using traffic flow data from Xiangfu Road Bridge, Changsha, Hunan Province, covering 30 days in March 2024, and used it to predict traffic flow on March 31st. Results indicate that the proposed model outperforms independent LSTM and Transformer models, achieving a prediction error of approximately 10%. Improved prediction accuracy supports intelligent transportation systems, enhancing real-time traffic management and infrastructure planning. The integration of MSA and LSTM significantly improves model robustness and provides a nuanced understanding of traffic dynamics, making it valuable for smart city applications.