Stock Image

A new artificial intelligence system developed by researchers at New York University could help cities forecast future traffic congestion before it becomes a major problem. The platform combines traffic forecasting, geographic mapping, and a locally hosted language model to give urban planners a way to identify emerging congestion patterns and explore potential changes.

The researchers tested the system using 15 years of publicly available New York City transportation data, covering observations from 2009 through 2024. The project was led by Anton Rozhkov, with Pranav Nitin Motarwar handling forecasting and Rudra Patil leading the spatial analysis.

The team compared two forecasting approaches: ARIMA, a conventional statistical model, and a Long Short-Term Memory (LSTM) neural network designed to identify complex patterns in time-series data.

When tested against previously unseen traffic data from 2021 through 2024, the LSTM produced a lower root mean square error than ARIMA. Its error was approximately 343 vehicles per day, compared with about 418 for ARIMA, representing an improvement of roughly 18%.

Using the LSTM model, the researchers projected average daily traffic could increase from approximately 12,540 vehicles in 2025 to 19,680 by 2029. They emphasized that these figures are forecasts rather than guarantees and carry inherent uncertainty.

The system also maps traffic patterns at a much more detailed geographic level. Researchers used Uber’s H3 geospatial system to divide the city into nested hexagonal areas, allowing congestion to be analyzed across boroughs and individual hotspots.

Clustering techniques identified locations where heavy traffic repeatedly appeared over several years. Manhattan recorded the highest overall congestion, followed by Brooklyn and Queens.

The platform adds another layer through Meta’s LLaMA language model. Planners can interact with a local knowledge base containing the project’s traffic forecasts, maps, and underlying data, allowing them to ask questions about potential infrastructure changes without manually examining large datasets.

The researchers deliberately designed the system to operate locally. This could allow government agencies to keep sensitive transportation information behind their own firewalls instead of sending it to external AI services.

However, the platform is not intended to make planning decisions independently, and it has not yet undergone a large-scale trial with working municipal planning departments. The researchers hope to test future versions with smaller municipalities and additional datasets.

The study demonstrates how forecasting, spatial analysis, and generative AI could be combined into a single tool for urban planning. Rather than replacing planners, the researchers envision AI helping them interpret complex transportation data and evaluate possible scenarios.

The research appears in a special issue of Transactions in GIS focused on ethical and explainable GeoAI, highlighting the importance of keeping the data, assumptions, and decision-making process transparent and under local control.