Spatio-Temporal AI Modeling for Urban Traffic Calibration: A SUMO-Based Approach

Abstract

Urban traffic management is a critical challenge in modern cities, necessitating innovative solutions to optimize traffic flow and reduce congestion. This research presents the development of an AI engine leveraging spatio-temporal learning techniques for urban traffic calibration. The proposed methodology leverages digital twin scenarios driven by microscopic simulations, which capture detailed vehicle behaviors—including interactions, lane changes, and driver dynamics to provide granular insights into urban traffic patterns. At the core of the AI engine is the Dynamic Spatio-Temporal Graph Attention Network (DSTGAT), a hybrid model that combines multi-head Graph Attention Networks (GATv2) with Long Short-Term Memory (LSTM) networks. DSTGAT exploits the joint spatio-temporal relationships inherent in traffic data by processing sequential snapshots of urban traffic, where each snapshot is represented as a graph with nodes indicating urban zones and edges carrying continuous flow values. The GATv2 layers, enhanced with residual connections and batch normalization, extract robust spatial embeddings, while the LSTM aggregates these embeddings over time to capture dynamic patterns and predict future traffic flows in real-time. The AI engine incorporates an iterative feedback loop that continuously refines the OD demand using synthetic scenarios, improving estimation accuracy across diverse urban environments. Preliminary results show that the DSTGAT‑based framework lowers OD‑estimation error on simulated data, suggesting its usefulness as an input to downstream traffic‑management strategies.