Multi-Task iTransformer: A Saturation-Based Model for Short-Term Highway Traffic Congestion Prediction Considering Event-Induced Capacity Variability

Accurate traffic congestion prediction is crucial for intelligent transportation systems (ITS), especially when accounting for the impact of random traffic events. Such events can lead to lane closures and variable road capacity, further complicating congestion prediction for traditional models. This paper proposes a novel multi-task iTransformer model designed specifically for short-term highway congestion prediction with consideration for random event impacts. Integrating an inverted embedding structure and multi-channel attention mechanism within a multi-task learning framework, the model simultaneously predicts traffic flow and capacity changes induced by events. Within a unified architecture, iTransformer captures the complex interactions among traffic flow, lane occupancy, and incident disturbances, enabling precise responses to dynamic traffic conditions. Empirical analyses based on the real-world data collected from highways in Zhejiang Province are conducted to validate the proposed model. It is shown that the proposed Multi-task iTransformer model performs well in traffic congestion prediction, achieving a 35.2% improvement in accuracy compared to single-task congestion prediction models.