Intersection-Level Turning Movement Flow Prediction Using an Adaptive Spatiotemporal Feature Fusion Network

Accurate and timely prediction of intersection-level turning movement flow (TMF) is critical for intelligent traffic signal control and real-time signal optimization. However, due to the high nonlinearity, complex interactions, and dynamic temporal variability of intersection traffic flow, the prediction task is more challenging than traditional road segment-level traffic flow forecasting. Currently, the majority of intersection-level TMF prediction studies predominantly adopt a single-model methodology for forecasting, often neglecting the potential contributions of exogenous features. To address this gap, this paper proposes a deep learning framework (AFF-GL) that parallelly integrates Residual Graph Convolutional Neural Networks (ResGCN) and Convolutional Neural Networks (ResCNN), with a two-layer Residual Long Short-Term Memory Network (ResLSTM). Firstly, in this framework, an improved ResGCN is designed to extract multi-scale features. Secondly, the ResCNN is constructed to capture time features. Then, an adaptive feature fusion mechanism is proposed to integrate the extracted information from the temporal and spatial relationship inherent in the traffic flow. Additionally, the incorporation of a two-layer ResLSTM further extracts the fusional spatiotemporal features and achieves the final prediction. At last, experimental results conducted on a real-world dataset collected from 15 intersections in Milton, Ontario, Canada, demonstrate that AFF-GL outperforms several existing renowned models in terms of prediction accuracy, stability, and generalization ability.