Multi-Scale Spatio-Temporal Graph Neural Networks for Urban Traffic Flow Prediction

Abstract

Accurate urban traffic flow prediction is crucial for intelligent transportation systems, urban planning, and resource allocation. Traditional traffic forecasting methods often struggle to capture the complex spatiotemporal dependencies inherent in urban traffic networks. This paper proposes a novel approach utilizing multi-scale spatio-temporal graph neural networks (MSST-GNNs) to address this challenge. The core idea is to integrate traffic data at various scales – street, regional, and city levels – and leverage the power of graph neural networks to learn intricate traffic patterns. The MSST-GNNs construct a graph representation of the urban road network, where nodes represent road segments and edges represent connectivity. The model then employs a multi-scale architecture to capture both spatial and temporal dependencies effectively. Specifically, we propose a hierarchical graph convolutional network (HGCN) that progressively aggregates information from finer to coarser scales, incorporating both spatial and temporal context. The HGCN is integrated within a recurrent neural network (RNN) to model temporal dynamics. The resulting MSST-GNNs achieve state-of-the-art performance on several benchmark urban traffic flow datasets. Experimental results demonstrate the effectiveness of the proposed approach in improving prediction accuracy compared to traditional methods and existing graph neural network models. The key contributions of this work are the novel MSST-GNN architecture, the hierarchical graph convolutional network, and the effective integration of multi-scale spatio-temporal data.