## TL;DR
AI is the operating system for urban mobility — predicting traffic flow hours in advance, optimizing traffic signals in real-time, and coordinating multi-modal transportation networks. With cities losing $300B+ annually to congestion, AI-driven intelligent transportation systems (ITS) offer the most immediate path to reclaiming lost time, fuel, and productivity.

## Core Explanation
Traffic flow prediction: given historical sensor data (induction loops, cameras, GPS probes) from road networks captured as time series, predict future traffic speed/volume/density at each sensor location. Time horizons: short-term (5-15 min, for signal control), medium-term (30-60 min, for route guidance), long-term (1-24 hours, for infrastructure planning). Spatial dependency: traffic at one intersection affects downstream intersections — the road network is a graph. Temporal dependency: morning/evening peaks, weekly patterns, holiday effects, incident disruptions. AI approaches: (1) Graph temporal models — DCRNN (Diffusion Convolutional RNN), STGCN (Spatio-Temporal GCN), Graph WaveNet — model spatial graph convolution + temporal dilated convolution; (2) Attention-based — ASTGCN, GMAN capture dynamic spatial dependencies (accidents change connectivity patterns) and long-range temporal dependencies via self-attention.

## Detailed Analysis
Nature 2025 ITS framework: three-stage pipeline — (1) Graph construction from road network topology with edge weights encoding distance and speed limit; (2) Spatial encoding via GCN aggregating neighboring sensor states; (3) Temporal encoding via transformer with positional encoding, capturing daily and weekly periodicity. Evaluated on PeMS (Caltrans Performance Measurement System, 35K sensors, 10 years of 5-min granularity data). LLM-based traffic optimization (ScienceDirect 2025): the LLM receives natural language descriptions of traffic state ("Northbound at intersection 42: queue length 15 vehicles, eastbound: 3 vehicles") and outputs signal phase decisions. The LLM's "common sense" reasoning about traffic dynamics complements the RL agent's data-driven optimization — the LLM catches edge cases (accidents, construction) that purely data-driven RL misses. MDPI 2025 xLSTM-based prediction: the recently-introduced xLSTM architecture (improved LSTM with exponential gating) achieves comparable accuracy to transformers at lower computational cost for real-time deployment. ScienceDirect 2026 advanced models review: fundamental challenge remains prediction under non-recurrent congestion (accidents, weather, special events) — these rare events have limited training data.

## Further Reading
- LibCity: Open-source traffic prediction library
- SUMO: Simulation of Urban MObility (DLR)
- PeMS: Caltrans Performance Measurement System