Researchers are combining artificial intelligence, geospatial mapping and clustering techniques to predict traffic congestion and help urban planners identify where interventions could have the greatest impact.
AI could help cities plan for less traffic congestion
Could artificial intelligence help determine how cities should be designed to reduce traffic congestion?
New York City has become a testing ground for a new GeoAI, or geospatial artificial intelligence, framework designed to forecast long-term traffic patterns, identify congestion hotspots and support data-driven urban planning decisions.
Two forecasting models put to the test
New York has an extensive database of transportation information. Researchers used 15 years of traffic observations, from 2009 through 2024, to test two methods for forecasting future congestion levels. The findings were published in a special issue of Transactions in GIS.
The two time-series forecasting models represent fundamentally different approaches. The first is ARIMA, a well-established statistical method. The second uses LSTM neural networks, a form of deep learning and modern artificial intelligence.
ARIMA, short for Autoregressive Integrated Moving Average, approaches forecasting by breaking the problem down into three distinct mathematical components. One of its main advantages is interpretability: its parameters make it possible to understand how much influence recent historical data has on a forecast.
Its structural limitation, however, is linearity. ARIMA assumes that relationships between data points over time can be represented through relatively simple linear equations. That can make it less effective when dealing with complex and dynamic phenomena or patterns influenced by unpredictable external variables.
LSTM, or Long Short-Term Memory, networks were developed to overcome some of these limitations by applying recurrent neural network architectures to sequential data.
In traditional neural networks, information may be processed without retaining a persistent memory of earlier inputs. LSTM networks, by contrast, can preserve relevant information across hundreds or even thousands of time steps, allowing them to identify nonlinear patterns, complex fluctuations and long-term dependencies.
The trade-off is that LSTM models typically require large amounts of data for effective training and can involve substantial computational costs.
“We started with an idea: What if we developed our own artificial intelligence platform that could be hosted locally and was secure, intuitive and convenient for planners to use in their daily work?” Rozhkov, one of the study’s three authors, said.
City governments are increasingly interested in integrating artificial intelligence into their operations, but planning departments often lack the technical expertise needed to build specialized AI systems. They may also be reluctant to upload sensitive transportation data to commercial AI models or chatbots.
AI traffic forecasting shows an 18% improvement
The study found that the LSTM model produced the strongest results when tested against previously unseen data from 2021 through 2024.
Its average forecasting error was approximately 343 vehicles per day, compared with about 418 vehicles for the ARIMA model, an improvement of roughly 18%.
The model then projected an increase in average daily traffic from 12,540 vehicles in 2025 to 19,680 in 2029.
Forecasting, however, was only one part of the project.
“A citywide average doesn’t help urban planners,” one of the researchers said. “What they need to know is which corridors are most affected, and those turn out to remain consistent year after year. That is exactly what the hexagonal mapping and cluster analysis were designed to explore.”
A planner could, for example, examine how a highway expansion near LaGuardia Airport might affect traffic in Queens, or what could happen if one-way streets in Lower Manhattan were converted into two-way streets.
Uber’s H3 system maps traffic patterns
In addition to forecasting vehicle volumes with the two models, the researchers implemented H3, an open-source geospatial indexing system developed by Uber that divides the Earth’s surface into a grid of similarly sized hexagonal cells.
Unlike square or rectangular grids, hexagonal cells provide more uniform relationships between neighboring cell centers. This can make calculations involving movement, distance and proximity more consistent across a geographic area.
Why clustering matters for traffic congestion
The researchers also used clustering, a machine learning and data analysis technique designed to identify groups of similar data points.
Instead of examining individual vehicles one at a time, the algorithm analyzes the hexagonal cells over time, looking for recurring patterns in which speeds decline or travel times increase.
This allows the system to distinguish between occasional congestion, such as traffic caused by an isolated accident, and systemic, recurring congestion, such as a major road that repeatedly becomes gridlocked during the evening rush hour.
By identifying where and when traffic repeatedly builds up, planners could make more targeted interventions, such as adjusting traffic-light timing, redesigning an intersection or roundabout, or expanding public transit services within a specific area.
According to the researchers, one of the project’s main strengths is its demonstration of how several emerging technologies can be combined into a single urban-planning workflow.
Rather than handing urban-planning decisions over to artificial intelligence, the researchers envision AI as an interface between planners and increasingly complex datasets, providing information that can help officials evaluate how changes to infrastructure and transportation systems could affect traffic over time.
