Making an Impact on Traffic and Energy Use: How the Collaboration with Uber is Moving our Research Forward
Anu Kuncheria, Graduate Student, Institute of Transportation Studies, UC Berkeley
Jane Macfarlane, Director Smart Cities, Institute of Transportation Studies, UC Berkeley and Lawrence Berkeley National Laboratory
Because of the significant computational power necessary to accurately simulate large-scale metropolitan transportation systems, city planners and transportation engineers must make do with reduced models, because the current tools used for simulation take hours or days to run and hence limit their ability to cover the scenarios of interest. This major roadblock could be solved by the simulation tool that we at the Institute of Transportation Studies at the University of California, Berkeley and Lawrence Berkeley National Laboratory have developed using high performance computation. Mobiliti, is a proof-of-concept, scalable transportation system simulator that implements parallel discrete event simulations on a supercomputer at Lawrence Berkeley National Laboratory. Our results have shown that it is possible to simulate one day of traffic dynamics in the Bay Area with 22 million vehicle trips, 100 percent dynamic rerouting - similar to your navigation app rerouting in traffic conditions - over a road network with 1.1 million nodes and 2.2 million links in less than 3 minutes of compute time!
Simulation, however, requires real-world validation to prove that the results generated are consistent and meaningful. Collecting data to be used as validation for sensing the real-world could be a challenging task. Uber is a rich source of robust data captured from vehicle location messages from over 17m trips a day across 700 cities. With its large fleet, Uber is constantly touching the infrastructure, logging large amounts of real time data. This data is highly valuable to research efforts focused on the best use of our road infrastructure today and will help build creative solutions for the future. Over the past 7 months, we have been directly using Uber Movement data to inform our simulations by validating travel times generated from the Mobiliti simulation.
This is a highly useful validation as Uber is a large participant in the transportation ecosystem. With the zone-to-zone level travel time comparison for trips, we were able to understand the mapping and demand anomalies in the simulator. For example, a group of travel time outliers were found to be caused by a poor distribution of origins and destinations across a geospatially large traffic analysis zone, Figure 1. As a consequence, we adjusted our algorithms to use a population-based distribution to rectify the problem and make the demand model more realistic.
Using the kepler.gl product we were able to track down anomalies for some other key traffic analysis zones, Figure 2. We are correcting these anomalies by integrating a professional map and improving the demand model further - provided by SFCTA - with a 40,000 micro analysis zone model and minute level temporal details. Figure 3 shows the Mobiliti generated speeds for the top 12,000 flow links in the Bay Area.
Figure 1: An automatically generated distribution of Origin/Destinations across a large TAZ that created unreasonable travel times. This was re-adjusted to account for population densities
Figure 2: Visualization of anomalies allows for tracking down inconsistencies in the complex simulation.
Figure 3: Visualization of speed data from Mobiliti on the 12K largest flow links
Going forward, we expect to also cross validate with the new Uber Speeds data. Speeds is a new dataset from Uber that provides historical aggregated speed data at the street segment level with coverage for all types of urban roadways.
Figure 4: Speeds dataset from Uber movement showing granular historical Speeds data in San Francisco
Our initial validation effort comparing Uber Speeds data and Mobility speeds compared across common links is shown in Figure 5 below. The links that show a difference greater than one standard deviation are shown in red and yellow. This will help us determine if our link models need to be adjusted to reflect congestion more accurately or if our demand models are not creating the congestion profiles that are typically seen on these links at this time of day.
Figure 5: Comparison of Uber and Mobiliti speeds for an 8am travel time. Green links are within one standard deviation, Red and Yellow links are lower and higher speed links respectively, that we plan to investigate further
The high-performance computational tools that we develop will help in the design of new active control ideas for future connected vehicles that will optimize energy, travel time and mobility for normal traffic conditions and networks under stress. Applications that can take advantage of this kind of simulation capability with this magnitude of reduction in computation time include:
- Emergency management planning where metropolitan scale simulations can be run for large numbers of scenarios;
- Freight Optimization where simulations can determine optimal routing to prevent queuing at Ports of Entry and opportunities for platooning;
- Traffic Management where opportunities for new mode shifts can be analyzed for mobility and energy optimization.
Our ongoing collaboration with Uber will allow us to begin to understand the very complex dynamics that we experience on our roadways, with the goal of making our city traffic more evenly distributed, more economical and fuel efficient.
Collaborators at LBNL: Cy Chan, Bin Wang, John Bachan, and Brian Gerke
The work described here is sponsored by the U.S. Department of Energy (DOE) Vehicle Technologies Office (VTO) under the Big Data Solutions for Mobility Program, an initiative of the Energy Efficient Mobility Systems (EEMS) Program. The following DOE Office of Energy Efficiency and Renewable Energy (EERE) managers played important roles in establishing the project concept, advancing implementation, and providing ongoing guidance: David Anderson and Prasad Gupte.