Highway Traffic Flow Prediction: Methods, Challenges, and Trends

Abstract

Traffic flow prediction is a core function of intelligent transportation systems. Accurate short‑term forecasts help reduce congestion, support signal control, and improve travel reliability. This paper reviews major methods for highway traffic flow prediction in clear, simple English. We organize the literature into four families: traditional statistical models, classic machine learning (ML), modern deep learning (DL), and hybrid approaches that combine multiple ideas. Statistical models (e.g., ARIMA) are simple and fast but struggle with nonlinear and spatial effects. ML methods (e.g., SVM, random forests, shallow neural networks) capture nonlinear patterns better but often require manual feature design. DL methods (e.g., LSTM/RNN, CNN, and graph neural networks) learn complex spatiotemporal patterns directly from data and now achieve state‑of‑the‑art accuracy on common benchmarks. Hybrids (e.g., ConvLSTM, graph‑attention networks, decomposition+DL) further improve robustness and accuracy. We also summarize widely used datasets and metrics, discuss open challenges (generalization, long‑horizon prediction, real‑time deployment, interpretability), and outline future directions (transfer learning, adaptive graphs and attention, uncertainty estimation, and cloud–edge deployment). The discussion is supported by real, citable studies and public datasets.