Real-time Traffic Congestion Prediction: A Novel Online Learning Method with Multi-Head Attention Mechanism and LSTM-Based Integrated Learning

Real-time traffic congestion prediction is essential for proactive traffic management, as it enhances the responsiveness of traffic systems, including route guidance, control, and enforcement. However, the heavy reliance on extensive historical data presents a significant challenge for real-time model updates. To overcome this limitation, this study proposes an advanced online learning framework that integrates a multi-head attention mechanism with LSTM-based ensemble learning. This approach incorporates traffic congestion factors as input features and employs average delay per kilometer as the predictive output. The experimental findings indicate that: 1) the proposed approach successfully enables real-time traffic congestion forecasting, and 2) it demonstrates strong adaptability in dynamic traffic environments.