Cycle-Based Traffic Flow Prediction of Signalized Urban Road Network

Traffic flow prediction is integral to urban intelligent transportation systems, especially for adaptive traffic signal control, where accurate cycle-level predictions are essential for minimizing delays and enabling green wave coordination. However, most existing models are designed for fixed-interval predictions on highways, ignoring the spatiotemporal complexities introduced by traffic signals in urban road networks. Specifically, traffic signals cause asynchronous flow sampling across intersections and discontinuous propagation patterns because of signal phase shifts, which severely limit the effectiveness of conventional prediction approaches. To address these challenges, this study introduces a novel deep neural network model, the adaptive graph convolution network, tailored for predicting cycle-based traffic flow. The model incorporates an adaptive calibration module (ACM) to align asynchronous and irregularly sampled traffic flow data across the signalized network, enhancing spatiotemporal adaptability in dynamic traffic environments. In addition, a message aggregation module (MAM) is developed to enable effective sharing of spatiotemporal information within urban signalized road networks. The efficacy of the proposed method is validated on a real-world urban road network controlled by adaptive signal timing. Experimental results show a reduction in the mean absolute error of 15.6% to 52.1% compared with existing prediction methods. Furthermore, the ablation experiment validated the effectiveness of the ACM and MAM in handling cycle-based traffic flow. These advances offer real time, signal-aware traffic predictions and support their practical application in green wave optimization.