PSR Research: Deep-learning traffic flow predictions

In Los Angeles, evening commutes take 81% longer than they would at other times; in the morning hours, commutes take 60% longer. These figures mean L.A. ranks as the most congested city in the United States and likely among the most congested cities in the world. For obvious reasons, this is a problem researchers, academics, politicians and Los Angeles residents are quite interested in solving. In their white paper “Deep-Learning Traffic Flow Prediction for Forecasting Performance Measurement of Public Transportation Systems,” Luan Tran, Min Mun, Yao Yi Chiang, and Cyrus Shahabi of the University of Southern California use the largest traffic sensor data warehouse in Southern California to better predict where and when this traffic occurs. 

First, they used deep learning techniques to create the Graph Convolutional Recurrent Neural Network (GCRNN), which takes data from tens of thousands of traffic sensors in the Southern California area and uses it to model traffic flow at different spatial and temporal resolutions. With this traffic prediction information in hand, they were able to build a bus estimated arrival time tool. Their hope with this tool was that more accurate predictions of bus arrival times would boost riders’ confidence in public transportation and help agencies coordinate their bus schedules, thereby helping move people out of their automobiles and into buses. 

Ultimately, their ETA model — which was then published to the web for public consumption — was 27% more accurate than the traditional prediction method, known as the Gradient Boosted Decision Tree. The researchers hope their ETA tool demonstrates how big data can help improve the performance and reliability of public transportation.

This research was supported by the Pacific Southwest Region University Transportation Center, the Region 9 University Transportation Center funded under the U.S. Department of Transportation’s University Transportation Centers Program. Established in 2016, the Pacific Southwest Region UTC is led by the University of Southern California and includes seven partners: Long Beach State University; UC Davis; UC Irvine; UCLA; University of Hawaii; Northern Arizona University; Pima Community College.

PSR conducts an integrated, multidisciplinary program of research, education and technology transfer aimed at improving the mobility of people and goods throughout the region. Its program is organized around four themes: 1) technology to address transportation problems and improve mobility; 2) improving mobility for vulnerable populations; 3) improving resilience and protecting the environment; and 4) managing mobility in high growth areas.