Transportation
Transportation accounts for 30% of energy consumption world-wide, and traffic delays are at an all-time high. In the United States, the average commuter loses 63 hours per year due to traffic delays, with the most congested areas reaching up to 137 hours of delays per year; traffic also costs about 40,000 lives every year.
Increasing efficiency and safety can improve the lives of millions, and save the lives of thousands.
Chattanooga Digital Twin (CTwin)
The goal of this 4-year project was to build a digital twin of the Chattanooga metropolitan area, with the ambitious ideal of reducing traffic-related energy use by 20%. The three primary steps in this effort were obsevability, predictability, and control. The project involved two national laboratories, two state departments of transportation, one city department of transportation, and two universities.
- Led the data science team in deriving insights from different sensor data through analysis, visualization, and machine learning.
- Served as technical lead during the final year of the project.
- Communicated with stakeholders from city and state departments of transportation and other organizations to understand their priorities, inquire details about the data provided, and share insights and results.
Observability: Situational Awareness
Achieving observability at a regional scale required the collection, fusion, and understanding of countless data sources. The team built CTwin, a situational awareness platform based on the EAGLE-I platform. This tool included a landing page dashboard to provide a quick overview of the system state, a map with over 30 layers. The map had interactive elements to display historic data for over 300 sensors and live data for over 200 sensors, which included radar sensors and CCTV video along the major highways, and 130 traffic signal sensors. CTwin also featured a page to compare traffic signals for selected corridors (defined via geojson polygon), a page to analyze Hamilton county 911 data (daily updates), and a page to analyze accidents at in detail (causes like lighting conditions and weather) from a E-TRIMS, a state-wide data source (weekly updates).
- Identified and collected over 120 different data sources, assessed their suitability for this project, and compiled a prioritized list of about 30 sources for further use.
- Majorly contributed to the development of CTwin, a real-time decision support platform which integrated data from over 30 sources and over 300 sensors (Docker, PostgreSQL, Angular, GeoServer, OpenLayers).
- Gained a deep understanding of GridSmart traffic sensors, their inner workings, and their API. This allowed me to propose more fine-tuned sensor configuration for more meaningful data collection.
- Developed a new way to visualize turn movements at intersections which was well-received by stakeholders.
- Mentored a postdoc on development of traffic emulation based on stationary sensor data.
- Significant Event Award for the successful deployment of CTwin 1.1
Prediction: Accident Detection
The project included work on simulating local traffic in SUMO, evaluating signal control methods, and near-term predictions of traffic volumes. My primary involvement in these aspects of the project was through data preparation.
My personal mission focused on traffic safety. After some previous analysis of accident statistics, and gaining a better understanding of the available data along the highway system, I began to wonder: can we identify traffic accidents in the sensor data?
A quick exploratorive analysis of traffic around the time of an accident indicated that yes – that should be possible. After implementing an inital prototype to validate that initial impression, I assembled a small team to carry out the broader vision.
- Formulated a novel methodology for accident detection on highway systems.
- Implemented a proof-of-concept using pure signal processing.
- Led the development of a machine learning-based accident detection tool which identified accidents 2-3 minutes sooner than reports via 911 calls.
Traffic Signal Control
Beginning at the end of the second year of the project, we began field experiments in the city we partnered with throughout this project: Chattanooga, Tennessee. I continue to be amazed that they actually let us do this! It really speaks to the great relationship between the city and the project team, and of course the great work of everyone involved in making our ambitious goals a reality. Throughout the project, there were two main control strategies which were suited to different testbeds.
- Shallowford Road: This is an arterial (all traffic lights in a row) which connects I-75 and Lee Highway to a major shopping district. There are 7 controllable traffic signals on this 1-mile stretch of road. At the time of our experiments, 6 of these intersections were equipped with GridSmart sensors which provided the required input data for our model. After several test sessions, we performed field testing for the Model Predictive Control algorithm during peak lunch hours (11:30 a.m.-1:30 p.m.) for 4 consecutive days in 2021. These experiments took place in live sessions with close monitoring of traffic, including queue spill-back on the highway ramps. We later ran the control 24/7 for three consecutive weeks with minimal supervision in 2022.
- Downtown Chattanooga: Downtown Chattanooga has a complex grid of connected traffic signals. Our test focused on 8 of these signals, which were located on 3 different roads. We performed field experiments for the Linear Quadradic Regulator signal control during morning rush hour (7:00-9:00 a.m.) during two full work weeks (i.e. Monday to Friday) 2022.
- Led the reconfiguration of GridSmart traffic sensors to collect all required data for control.
- Orchestrated experiments for a real-time traffic signal control in close collaboration with NREL and CDOT, and achieved a 19.4% travel time reduction compared to time-of-day scheduling.
- Shallowford Road: Model Predictive Traffic Signal Control
- Downtown: Traffic Signal Control using Linear Quadratic Regulators
Real-Time Control for Connected Autonomous Vehicle (CAV) Ecodriving
This project was a collaboration with Toyota and the City of Chattanooga. The goal of this project was to develop ecodriving for autonomous vehicles. The main premise is this: when you drive down the road, you use a lot of energy to get to your maximum speed, but from there, you can coast with relatively little energy (assuming the road is flat). When you brake (e.g. to stop at a traffic light), you then have to spend all that energy again to get back to your previous speed. However, if you already knew that the traffic signal you’re approaching will turn green in a certain number of seconds, you might be able to slow down enough to delay your arrival until the signal is green, and coast through instead of stopping.
To share this information with vehicles, we needed a way for cars and traffic signals to exchange information (Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communication). We decided to implement this using a subscription-based system. The vehicle would then automatically subscribe to the right channels (nearby traffic signals which were downstream from the current location). There was a caveat to this: regulations didn’t allow us to actually let the car drive in real traffic on its own. To compensate for that, we developed a mobile app which would connect to the system and show the speed recommendations to a human driver.
We performed labratory testing of the ecodriving algorithm on a vehicle dynamometer (a device that attaches to the vehicles axles and measures movement) which was set up in front of a screen showing VisSIM real-time simulations, and the driver followed the mobile app recommendations. In 2023, we performed field testing of the speed control algorithm, combined with traffic signal control, on Shallowford Road in Chattanooga, Tennessee.
- Coordinated cross-functional team, including industry partners and researchers, to ensure smooth integration of vehicle on-board software, cloud-based communication system, and traffic infrastructure.
- Developed subscription-based V2I and V2V communication between connected and autonomous vehicles (Linux, channels), traffic infrastructure (REST APIs), and a web application (OpenStack, django).
- Derived insights from 6 GridSmart traffic sensors and on-board vehicle sensors through data analysis and visualization.
- Designed and developed data-driven coupling between historic and real-time sensor data, and coordinated integration with predictions, VisSIM simulations, and visual analytics components.
- Supported on-road field testing through remote monitoring.
- Best Poster and Demo Paper for our IEEE MASS 2022 paper Mobile App for Ecodriving
Campus Digital Twin
This project was part of the Sustainable ORNL Initiative. The goal was to build a smart and sustainable traffic system which would inform employees of traffic flow, available parking spots, and available electric vehicle charging infrastructure.
- Predicted origins and destinations of employee vehicles based on traffic data at traffic signals (GridSmart), campus gates (CCTV converted to count data using YOLO), and parking lots (CCTV+YOLO).
- Supported the analyis of electric vehicle charging data.
