Deep Learning Approaches for Traffic Forecasting: Spatiotemporal Heterogeneity and Model Advances

Sichuan University, Chengdu, Sichuan, China

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

With the increasing complexity of urban traffic systems, accurate traffic flow forecasting has become a critical component of intelligent transportation systems. Traditional statistical and machine learning models, such as Autoregressive Integrated Moving Average (ARIMA) and random forests, struggle to address the non-stationarity and heterogeneity of traffic data. In recent years, deep learning has significantly advanced spatiotemporal traffic forecasting. This paper provides a comprehensive review of deep learning-based models, including Diffusion Convolutional Recurrent Neural Network (DCRNN), Graph WaveNet, Spatio-Temporal Decoupling Masked Autoencoder (STD-MAE), and Spatio-Temporal Pivot Graph Neural Network (STPGNN), analyzing their structural designs and performance improvements. By examining the theoretical foundations of temporal and spatial heterogeneity, this study highlights how deep models effectively capture multi-scale dependencies, dynamic topologies, and pivotal node behaviors. The paper also discusses challenges such as data noise, computational cost, and real-world deployment. Future research directions include transferable pre-training, frequency-domain modeling, and lightweight architectures to achieve efficient and interpretable traffic forecasting.