Traffic forecasting for every city?
Connecting state and local government leaders
Using a new transfer learning approach, researchers were able to analyze traffic data from the San Francisco region and forecast traffic flow in the Los Angeles region and vice versa.
Researchers from Argonne National Laboratory are using a graph neural network and a new transfer learning approach to more accurately forecast large-scale traffic patterns.
Using historical data from sensors installed along California highways, the researchers used a diffusion convolutional recurrent neural network to train a model to forecast traffic speed and flow. Within milliseconds, the model can analyze the past hour of data and accurately predict the next hour’s traffic patterns, Argonne officials said.
By breaking the large traffic network into a number of smaller ones and training each one separately, the researchers were able to greatly increase the speed and effectiveness of the overall model. The new model’s predictions have been shown to be within six miles per hour of the observed speed at any location.
What’s more, a model trained on “data-rich regions of highway network can be used to forecast traffic on unseen regions of the highway network,” the researchers wrote in their paper. Using a new transfer learning approach, the model was able to analyze traffic data from the San Francisco region and forecast traffic flow in the Los Angeles region and vice versa, they said.
Accurate traffic forecasting is an especially thorny problem because it has both complex spatial and temporal dependencies. The supercomputing resources at Argonne drastically reduced the amount of time needed to train the model that pulled data from a year’s worth of sensor readings from over 11,000 locations.
“The scale of this project is large, and this amount of data requires an equally large computing resource to tackle it,” said Prasanna Balaprakash, a computer scientist with the Mathematics and Computer Science division and the Argonne Leadership Computing Facility. “The AI and supercomputing capabilities that have been used in this work allow us to tackle really large problems.”
“Data-driven forecasting methods can help better utilize the transportation resources, develop policies for better traffic control, optimize routes, and reduce incidents and accidents on the road,” the researchers said. Insights gained from the analysis can also inform intelligent transportation systems, dynamic routing for freight traffic and congestion pricing for toll roads.
“These methods can allow states, cities, and municipalities to more quickly develop improved traffic forecasting capabilities with a significantly smaller infrastructure investment,” they said.
NEXT STORY: Study: The Best—and Worst—State Highway Systems