Research Projects

Stop the Video

Research Projects

STATUS: In Progress YEAR: 2020 TOPIC AREA: Public transit, land use, and urban mobility Transportation planning, policy, and finance CENTER: PSR

Large-scale and Long-term Forecasting of Performance Measurement of Public Transportation Systems

Project Summary

Funding source: Caltrans
Contract number: 69A3551747109
Funding amount: $100,000
Start and end dates: 08/16/20 to 08/15/21

Project description

The main objective of this proposal is to develop a machine learning approach for large spatial scale (e.g., the entire Los Angeles Metropolitan Area) and long-term (e.g., days instead of hours) performance measurement forecasting for public transportation systems. Reliable long-term forecasting of performance measurement for public transportation systems over a large area is essential for policymakers to achieve effective city planning as well as promotes ridership. For example, forecasting bus arrival time for the next day helps a rider to plan their commute early. Existing approaches typically rely on traffic simulation tools and models that require expert knowledge to execute and adjust parameters for various traffic scenarios. Previously under a Caltrans funded project, we developed a data- driven, deep learning approach for traffic flow prediction and bus arrival time estimation. We also built a system that uses the traffic predictions for forecasting various performance metrics for public transportations in Los Angeles (e.g., bus arrival times). The system demonstrates the overall approach in an area near downtown Los Angeles and shows that incorporating traffic flow predictions can help to forecast short-term bus arrival times accurately (e.g., in the next few hours).

In this proposal, we plan to build on our current approach and system to develop the capability for processing the entire Los Angeles Metropolitan Area (and extendable to the entire state) for long-term forecasting (e.g., days instead of hours) of a variety of public transportation system performance metrics. The major challenges include 1) how to predict/infer traffic speed on a road segment when there is no traffic sensor at the road segment, 2) how to overcome the computational and memory bottlenecks of training a deep learning network for traffic flow forecasting over a large road network, and 3) how to deal with the large error propagation in the deep learning network for forecasting over a long-time span. We will explore both spatial statistical methods and machine learning methods for the estimation of traffic speed for the road segments that do not have a traffic sensor. We will investigate methods to represent large road networks into a condensed graph or multiple subgraphs so that the deep learning network can scale to process large spatial areas. We will also study various training methods and network architectures (e.g., teacher forcing) to enable long-term forecasting of both traffic flows and the performance measurement of public transportation systems.

To demonstrate the results from the proposed research, we will develop a web application in which users can select the start time of a trip and access the predicted bus arrival times anywhere in the entire Los Angeles Metropolitan Area. This research will exploit real-world big traffic sensor data and California Highway Patrol (CHP) accident logs collected from the Regional Integration of Intelligent Transportation Systems (RIITS) in the last nine years under our existing Archived Traffic Data Management System (ADMS) project.


Cyrus Shahabi
Helen N. and Emmett H. Jones Professor of Engineering
3737 Watts Way
Charles Lee Powell Hall (PHE) 306ALos Angeles, CA 90089-0781
United States
[email protected]