Optimization Study of Short-Term Traffic Flow Prediction Model Based on Spatiotemporal Graph Convolutional Network

Abstract

Short term traffic flow prediction is one of the basic problems of intelligent transportation systems (ITS) which directly affects urban mobility management, route planning and congestion mitigation. The existing prediction methods such as statistical time-series models and shallow machine learning methods do not effectively and simultaneously reflect the spatial dependence between road sensors and the non-linear temporal evolution of traffic. Current approaches mostly use fixed, pre-defined road network adjacency matrices, which fail to capture the dynamic nature of real-world road traffic interactions, and achieve good performance by deep learning algorithms, such as Graph Convolutional Networks (GCNs) and temporal sequence models. In this paper, we propose a novel framework called Adaptive Spatiotemporal Graph Convolutional Network with Multi-Scale Attention (ASTGCN-MSA) to learn the spatiotemporal dependency jointly from traffic sensor data. Three major innovations are introduced in the proposed model: (1) Adaptive Graph Learning (AGL) module that incorporates the prior road topology and leverages the data-driven learned adjacency matrix on the graph, using trainable node embeddings; (2) Multi-Scale Temporal Attention (MSTA) module that applies parallel dilated 1-D convolutions with three kernel sizes (k = 1, 3, 5) over the temporal dimension of the graph with channel-wise sigmoid gating; and (3) Spatial Self-Attention (SSA) module that uses multi-head attention over node embeddings to capture long-range inter-sensor dependencies beyond the local graph neighborhood. The model is tested using the Traffic Flow Forecasting benchmark data set, consisting of 36 sensor sites on major highways in the Northern Virginia/Washington D.C. capital region, and sampled at 15-minute intervals. These extensive experiments show that ASTGCN-MSA consistently performs best across all metrics under comparison: MAE, RMSE, MAPE and R², compared to four competitive baseline models: LSTM , GRU , Vanilla STGC and Transformer . The individual impact of each suggested component is verified through ablation studies, while hyperparameter sensitivity analysis offers practical tips for model setup. Results demonstrate the ability of ASTGCN-MSA to predict the traffic flow in the real world; it is scalable, robust, and interpretable. This paper introduces a novel approach that leverages graph convolutional networks and spatiotemporal learning to predict traffic flow at an early stage.A novel approach is developed by introducing the concept of graph convolutional networks and spatiotemporal learning to predict traffic flow at an early stage.