A New AI Framework Could Help Cities Plan for Future Traffic

Traffic jams are usually treated as a problem to be measured after they form. Anton Rozhkov and his colleagues are exploring a different approach: giving planners state-of-the-art tools that can anticipate congestion, locate where it is most likely to emerge and translate complex traffic models into plain-language guidance.

In a new study published in the special issue of Transactions in GIS on Ethical and Explainable GeoAI, Rozhkov, Pranav Nitin Motarwar and Rudra Patil, all affiliated with NYU Tandon, developed a geospatial artificial intelligence (GeoAI) framework that combines traffic forecasting, spatial mapping and a locally hosted large language model. Rozhkov led the project and designed the framework, with Motarwar leading the forecasting work and Patil the spatial analysis. The system was tested using 15 years of New York City traffic data.

The motivation came partly from conversations Rozhkov had with urban planning practitioners. Cities increasingly want to incorporate AI into their work, he says, but planning departments may lack the technical expertise to build specialized systems. They may also be reluctant to upload sensitive transportation data to commercial models and chatbots.

“We started with an idea: what if we developed our own AI platform, one that could be hosted locally and would be secure, intuitive, and comfortable for planners to use in their day-to-day work,” Rozhkov said, describing a system that could be hosted locally and used directly by planners.

Traffic offered a useful test case because New York makes extensive transportation data publicly available. The researchers analyzed traffic observations from 2009 through 2024 and tested two methods for predicting future volumes. One was ARIMA, a conventional statistical forecasting technique. The other was a Long Short-Term Memory, or LSTM, neural network, which can identify nonlinear patterns extending across long sequences of data.

Motarwar, who trained the forecasting models, benchmarked the two approaches against each other. “ARIMA gives you seasonality, which is true but not the whole story,” he said. “The LSTM picks up the parts of the pattern that don’t repeat cleanly, and that’s where most of the improvement came from.”

On previously unseen data from 2021 through 2024, the LSTM performed better. Its average prediction error was about 343 vehicles per day, compared with roughly 418 for ARIMA, an improvement of approximately 18 percent. The model then projected average daily traffic increasing from 12,540 vehicles in 2025 to 19,680 in 2029, although those figures come with substantial prediction intervals and should be understood as forecasts rather than certainties.

But forecasting was only one piece of the project. Patil led the spatial component, mapping traffic onto Uber’s H3 system, which divides geographic space into nested hexagonal cells. Unlike citywide averages, this approach allows planners to examine congestion at multiple scales, from borough-level patterns down to localized hotspots. Clustering algorithms then identified areas where heavy traffic repeatedly accumulated. Unsurprisingly, Manhattan emerged as the highest-congestion borough in the analysis, followed by Brooklyn and Queens.

“A citywide average doesn’t help the city planners,” Patil said. “What they need to know is which corridors are badly impacted, and those turn out to be consistent year to year. That is what the hexagonal mapping and clustering were designed to explore.”

The final component turns those analytical results into something closer to a conversation. Building on the forecast and congestion maps, the researchers built a customized portal using Meta’s LLaMA model connected to a project-specific knowledge base containing forecasts, congestion locations and other traffic information. A planner could ask, for example, how a highway expansion near LaGuardia Airport might affect Queens traffic or what could happen if one-way streets in Lower Manhattan were converted to two-way streets.

Crucially, the system is designed to run locally. Agencies could keep sensitive transportation information “behind their own firewall,” Rozhkov said, rather than sending it to an outside AI service, even if sometimes the capabilities of the commercial models might be better. “A general chatbot gives you a reasonable-sounding generic paragraph about congestion,” Motarwar added. “Planners need an answer that comes from their own data. We saw this as much a data problem as a model problem.”

The researchers are not claiming that the platform can solve congestion on its own. Nor has it yet undergone a large-scale trial with working planners. Rozhkov hopes to test future versions with smaller municipalities and other planning datasets. Instead, the project demonstrates how several emerging tools can be combined into a single planning workflow. Rather than asking AI to make urban decisions, the researchers envision it as an interface between planners and increasingly complicated datasets.

That distinction also explains why the paper appeared in a special issue devoted to ethical and explainable GeoAI. For Rozhkov, the question is not simply whether cities can use increasingly powerful AI. It is how they can use it while keeping the underlying data, assumptions and decisions under local control that will eventually benefit local residents.