Graph enhanced spatial–temporal transformer for traffic flow forecasting

Abstract

Traffic flow forecasting, which aims to predict future traffic patterns based on current conditions, is a crucial yet challenging task in intelligent transportation systems due to the complex spatial–temporal relationships involved. Existing methods often struggle to effectively capture these intricate spatial dependencies and temporal patterns. To address these limitations, we propose a graph enhanced spatial–temporal Transformer (GE-STT), which integrates a graph enhanced module and a spatial–temporal Transformer module for improved prediction accuracy. Specifically, the graph enhanced module combines a Graph Convolutional Network (GCN) with a Gated Recurrent Unit (GRU) to obtain enriched spatial–temporal features, and introduce the original traffic data as a correction term to deal with the errors in the enhancement process. The spatial–temporal Transformer then leverages these enhanced features for final prediction. Experimental results on four traffic datasets show that GE-STT achieves superior performance under various metrics. Compared with the best baseline in different datasets, the performance of GE-STT is improved by up to 8% under the MAE metric, highlighting its robustness and effectiveness in traffic flow forecasting tasks.