Spatiotemporal AI for Intelligent Transportation Systems: Methods, Datasets, and Benchmarks

Abstract

The rapid growth of urban mobility and the increasing demand for sustainable traffic solutions have positioned Intelligent Transportation Systems (ITS) as a central research area. Spatiotemporal AI integrates temporal and spatial elements to enhance transportation systems performance and forecasts. This paper evaluates advances from 2020 to 2025 in four different approaches: Graph Neural Networks (GNNs), Transformer models, Reinforcement Learning/Multi-Agent RL, and their combinations using popular datasets (METR-LA, PEMS-BAY, LargeST) and standard evaluation metrics (MAE, RMSE, MAPE). According to the results, GNNs excel at processing spatial data, Transformers perform well with lengthy sequences, and RL techniques can improve the performance of adaptive controls. Combination models may be able to strike a good balance between accuracy and speed. The study also assesses models based on predictive accuracy, scalability, interpretability, and computational cost. However, real-time deployment, merging data from various sources, and making it work for entire cities still present challenges. To improve the next generation of intelligent street mobility, the study suggests looking into edge computing, federated learning, and physics-based AI.