Abstract
Operating a joint network of highways and arterial streets in real time is challenging. The main challenges are twofold. Highways and arterials are highly inter-dependent, but may have their own operational strategies and systems that do not necessarily synchronize. As a result, traffic queues can spillover from highway to arterials, or the other way around, leading substantial congestion that worsens the system performance. Coordinating the signal control system on arterials and ramp metering control on ramps are key to mitigating such congestion. In addition, most signal or ramp metering systems deal with recurrent traffic congestion or normal traffic conditions. They can alleviate queues locally to some extent under non-recurrent congestion (being responsive or reactive), but are not designed to prevent queuing from the occurrence of incidents (being predictive) nor mitigate congestion for the joint network. To this end, managing traffic predictively (or proactively) and coordinating ramp metering and street signals among all relevant highway on-ramps/off-ramps can effectively improve the joint network performance. Transportation Systems Management and Operations (TSMO) refers to a set of strategies that could be utilized to mitigate system-level congestion, particularly non-recurrent traffic impacts, such as information provision, signalization, and access control. Though TSMO are technically available to practitioners, but what time and what strategy to engage remain unknown. Being predictive and proactive, and coordinating among all control strategies (e.g. street signals and ramp metering jointly) is the key to effective management of network-level traffic. Proactive operational management is highly dependent on accurate real-time traffic data and swift real-time traffic prediction.
This research project addresses two problems for an integrated TSMO system: ahead-of-curve prediction and system-level signal and ramp metering coordination. We propose to develop theories, models and algorithms of machine learning to predict traffic patterns in real time being a typical recurrent pattern or non-recurrent pattern, and to optimize the timing plans for both ramp metering and street signals in the TSMO system. Prediction and operational strategies are intimately coupled. The prediction will be made by a machine that learns not only historical traffic patterns but also real-time data (possibly from multiple sources). Operational strategies are made and updated in real time to achieve management goals (e.g. minimization of total travel time) as a result of ahead-of-curve prediction of network impacts. In particular, the research will fuse multiple data sources related to highways and local street/intersections; develop an efficient network-level modeling framework enabled and validated by multi-source data; make real-time optimal signal plans and ramp metering plans; and finally quantify the network benefits of operational strategies to improve mobility/safety.
Description
Our general research approach is to develop the integrated corridor control into two stages. The first stage leverages large-scale multi-source traffic data (e.g. 5-min counts, speed by cars and trucks, etc.) to predict traffic conditions in 30-60 minutes advance, assuming current traffic control strategies remains unchanged. This requires using machine learning and network simulation techniques to detect system anomalies (or recurrent system conditions) in real time and predict traffic progression at the system level without engaging adaptive control strategies. The second stage proactively engages adaptive control schemes for both ramps and local intersections in response to predicted traffic conditions. Those control schemes are optimally selected to minimize system-level congestion while ensuring equity among highways and arterials. Those adaptive control schemes are developed offline and can be adjusted in real time adapting to real-time traffic. Those offline schemes are twofold: one developed for recurrent traffic conditions under a few typical patterns, while the other for non-recurrent traffic conditions. The main advantage of this two-stage corridor control method is to simplify the system dynamics within control logics as either a trained machine or offline simulation, which ensures efficient and timely implementation of control algorithms. It also works for both recurrent and non-recurrent traffic conditions that would require completely different tactics in terms of coordination among all on-ramps, off-ramps and adjacent intersections in the network. We would also try to implement a model-free reinforcement learning algorithm in the second stage. Its advantage over the model-based optimization algorithm is that it can be computationally efficient and adaptive to real-time traffic.
Task 1: Identify various data sources for in-depth data analytics and system control
1. Request GIS models (roadway infrastructure features, e.g., topology, number of lanes, intersection layout, speed limit, etc.) and regional travel planning model from MDOT where origin-destination (O-D) trip information associated in the region can be retrieved. The OD trip information is optional, but can improve the overall modeling accuracy.
2. Establish a refined GIS network model for this research in the TSMO 1 area.
3. Obtain traffic counts on local streets, intersections and highway in the TSMO 1 area from MDOT. To produce accurate modeling results, 15-min traffic counts are necessary but 1-min or 5-min intervals are preferred.
4. Obtain INRIX probe data from RITIS and HERE probe data from MDOT. Those probe data cover highways in the region and major arterials within the TSMO 1 area.
5. Obtain existing signal timing schemes and current ramp metering schemes from MDOT.
6. Work with MDOT and contractors to gather detailed information about the management goals in the TSMO 1 area, such as queue limits on on-ramp and off-ramps, as well as on local streets.
7. Work with MDOT and contractors to collect historical incident data, including geographical scope of the closures, lane closure configurations, crashes and past events that substantially influence traffic in TSMO 1 area.
Task 2: literature review
Conduct thorough literature review on coordinated ramp metering and arterial signal timing in the following aspects:
1) Real-time traffic state estimation in the joint network of highways and arterials, typically 5-10 min in advance.
2) Real-time traffic flow prediction in the joint network of highways and arterials, typically 5-60 min in advance.
3) Coordinated ramp metering and arterial signal timing (also known as local signal synchronization): rule-based control method. This includes deterministic optimization of metering rates and signal timings given each fixed traffic flow pattern, which provides deterministic rules on coordinated control schemes in response to a specific traffic flow pattern.
8. Coordinated ramp metering and arterial signal timing (also known as local signal synchronization): AI-based control method. This includes reinforcement learning techniques that allow for real-time tuning of metering rates and signal timings, accounting for futuristic demand uncertainty.
Task 3: Develop a dynamic network model for the TSMO 1 system
In this task, will develop a dynamic network model for TSMO 1 that provides estimated 5-min origin-destination demand among all street segments that vary by time of day. The route choices for all travelers in the area will be examined and carefully calibrated using multi-day data sets collected in Task 1. The network model is capable of estimating network-wide traffic impact caused by 1) any incident in the TSMO network consisting of freeway and major arterials; 2) any signal timing; 3) any ramp metering rates. It has the capacity of modeling dynamic traffic evolution with the consideration of real-time travel control and traffic demand management. It adopts state-of-the-art traffic models and is much more computationally efficient than other microscopic models that are extremely labor intensive to build.
The TSMO network, together with some peripheral areas of TSMO, will be coded into the dynamic network model. Baseline travel demand will be estimated in the first place using the integrated traffic data (counts, INRIX data by cars and trucks) on typical weekdays without the presence of incidents. In addition, the overall traffic impact for historical incidents without engaging corridor control can be measured by time-of-day traffic evolutions in the area, as well as performance metrics, such as total traffic delay, average travel time, emissions, energy use, vehicle-miles traveled, etc. Those performance metrics will be estimated for both highways and arterials, respective to ensure equitable allocation of road capacities.
In addition, we train a sophisticated machine with large-scale multi-source traffic data (e.g. 5-min counts, speed by cars and trucks, etc.) to predict traffic conditions in 30-60 minutes advance, assuming current traffic control strategies remains unchanged and without engaging corridor control. This requires using machine learning and network simulation techniques to detect system anomalies (or recurrent system conditions) in real time and predict traffic progression for each road segment. We will develop a collection of statistical models to analyze the statistical differences of TSMO network flow among days. The objective is to rigorously distinguish all typical recurrent network flow patterns (e.g., day of week or seasonal) and non-recurrent flow patterns. The machine learning based prediction allows us to accurately forecast the traffic conditions in at least 30 min ahead under two cases: 1) recurrent traffic; 2) those incidents that occurred historically. For any traffic prediction with incidents that do not occur in the past, the dynamic network model needs to be employed to make the forecast.
The dynamic network model and its data-driven framework will be developed based on Mobility Data Analytics Center - Prediction, Optimization, and Simulation Toolkit for transportation Systems (MAC-POSTS) led by the CMU team. MAC-POSTS is a mesoscopic traffic simulation software empowered by state-of-the-art machine learning and network flow algorithms.
Task 4: Develop control strategies for ramp metering and local signal synchronization
In this task, metering rates at different meters along the corridor and related arterials signals are optimally identified to minimize system-level congestion while ensuring equity among highways and arterials, for each typical recurrent patterns and non-recurrent patterns. Those optimal control schemes for each traffic pattern are determined offline and can be adjusted in real time adapting to real-time traffic. For each traffic pattern, we will use network optimization techniques (examples include Chang et. al, 2020) to solve for the optimal timings and metering rates. Scenarios of various travel demand with or without incidents will be created and modeled based on the network model under Task 3. We will then test the two-stage coordinated control algorithms under those scenarios. In particular, rules or strategies that select an optimal scheme in real time will be determined under the network simulation environment.
Task 5: Evaluate effectiveness of optimal corridor control for the TSMO 1 system
This task evaluates the TSMO system performance (total delay on both highways and arterials) before and after the deployment of corridor control under those scenarios. This will be completed in a simulation environment, but can serve as a benchmark of control system performance before field deployment in the future. One key component is to use actually observed traffic data to define the system performance metrics that may vary from the simulation environment. We will work closely with MDOT SHA to identify system performance metrics for TSMO 1. The initial performance will be measured by at least two metrics: 30-min-ahead prediction accuracy of traffic flow in the TSMO 1 area; and reduction in total traffic delay in the TSMO 1 area and its proximity area (comparing to historical data under a similar traffic condition).
Timeline
Task 1 and kick-off meeting: one month
Task 2: one month
Task 3: four months
Task 4: four months
Task 5: six months
Strategic Description / RD&T
Deployment Plan
We will work closely with the Office of the Transportation Mobility and Operations (OTMO) (formerly the Office of CHART and ITS Development) at MDOT to implement this research results. The team consisting with Morgan State University (MSU) and Carnegie Mellon University (CMU, subcontractor) will hold a joint MSU-CMU bi-weekly coordination calls to discuss difficulties encountered and proposed solutions, and to outline plans for completing the scope of work, key milestones and deliverables. When performing the tasks, we will together meet with OTMO project managers, engineers and TSMO staff who provides feedback/comments for each quarter, to ensure the model development and testing are consistent with MDOT’s view, and the tasks are aligned with the partners’ needs.
In terms of implementation barriers, we will evaluate and prioritize barriers as the project progresses. Main potential barriers for this project are to develop efficient and robust solution algorithms for data-driven network models and optimization of control strategies. If any risks or barriers are identified during the project, we will use our domain expertise to find alternative methods or to seek professional help from both data and methodology perspectives utilizing resources from both MSU and CMU. The core MSU-CMU team has sufficient and somewhat overlapping expertise that we can reallocate personnel if needed. Another advantage is that this research team have been intensively collaborating on research projects in the past two years with necessary optimization and traffic engineering skill sets, and thus are able to reduce or eliminated those barriers.
This project will be managed by MDOT SHA. CMU-MSU and MSDOT will have monthly meeting to track the progress of the project. We will receive feedbacks from MDOT and test the method and system in the TSMO 1 system.
Expected Outcomes/Impacts
We anticipate to develop a software package for dynamic network modeling and data analytics The software package would specifically address real-time prediction, and real-time coordinated corridor control, using TSMO 1 system as demonstrative case studies.
This project evaluates the TSMO system performance (total delay on both highways and arterials) before and after the deployment of corridor control under those scenarios. This will be completed in a simulation environment, but can serve as a benchmark of control system performance before field deployment in the future. One key component is to use actually observed traffic data to define the system performance metrics that may vary from the simulation environment. We will work closely with MDOT SHA to identify system performance metrics for TSMO 1. The initial performance will be measured by at least two metrics: 30-min-ahead prediction accuracy of traffic flow in the TSMO 1 area; and reduction in total traffic delay in the TSMO 1 area and its proximity area (comparing to historical data under a similar traffic condition).
Expected Outputs
TRID
Individuals Involved
| Email |
Name |
Affiliation |
Role |
Position |
| sbenicky@andrew.cmu.edu |
Benicky, Sheryl |
CIT ERA |
Other |
Other |
| seanqian@cmu.edu |
Qian, Sean |
CEE/Heinz |
PI |
Faculty - Untenured, Tenure Track |
| weiran@cmu.edu |
Yao, Weiran |
CEE |
Other |
Student - PhD |
| qilingz@andrew.cmu.edu |
Zou, Qiling |
CEE |
Other |
Other |
Budget
Amount of UTC Funds Awarded
$50000.00
Total Project Budget (from all funding sources)
$120000.00
Documents
| Type |
Name |
Uploaded |
| Project Brief |
Qian_mdot_2021.pptx |
March 14, 2021, 8:58 p.m. |
| Data Management Plan |
dmp_Aky8dGf.docx |
March 14, 2021, 8:58 p.m. |
| Progress Report |
367_Progress_Report_2021-09-30 |
Sept. 23, 2021, 9:19 p.m. |
| Presentation |
Dynamic O-D estimation |
March 16, 2022, 3:44 p.m. |
| Progress Report |
367_Progress_Report_2022-03-30 |
March 16, 2022, 3:44 p.m. |
| Publication |
Accessibility and The Crowded Sidewalk: Micromobility's Impact on Public Space |
May 2, 2022, 9:26 a.m. |
| Publication |
Lane Management with Variable Lane Width and Model Calibration for Connected Automated Vehicles |
May 2, 2022, 9:27 a.m. |
| Publication |
SUSTAIN: Sustainable Urban Systems Through Automated Infrastructure Networks |
May 2, 2022, 9:27 a.m. |
| Publication |
High-resolution traffic sensing with autonomous vehicles |
May 2, 2022, 9:28 a.m. |
| Publication |
Optimized graph convolution recurrent neural network for traffic prediction |
May 2, 2022, 9:28 a.m. |
| Publication |
Lane management with variable lane width and model calibration for connected automated vehicles |
May 2, 2022, 9:29 a.m. |
| Publication |
Path-based system optimal dynamic traffic assignment: A subgradient approach |
May 2, 2022, 9:29 a.m. |
| Publication |
A low rank dynamic mode decomposition model for short-term traffic flow prediction |
May 2, 2022, 9:30 a.m. |
| Publication |
Learning to recommend signal plans under incidents with real-time traffic prediction |
May 2, 2022, 9:30 a.m. |
| Publication |
Statistical inference of travelers' route choice preferences with system-level data |
May 2, 2022, 9:31 a.m. |
| Publication |
Estimating probabilistic dynamic origin-destination demands using multi-day traffic data on computational graphs |
May 2, 2022, 9:32 a.m. |
| Publication |
Inferring the causal effect of work zones on crashes: Methodology and a case study |
May 2, 2022, 9:32 a.m. |
| Publication |
A novel map-matching algorithm for relating work zones and crashes |
May 2, 2022, 9:33 a.m. |
| Publication |
Hierarchical graph convolution networks for traffic forecasting |
May 2, 2022, 9:33 a.m. |
| Publication |
High-resolution traffic sensing with probe autonomous vehicles: A data-driven approach |
May 2, 2022, 9:34 a.m. |
| Presentation |
Dynamic O-D estimation |
Sept. 29, 2022, 8:02 p.m. |
| Presentation |
TSMO 1 system simulation with ramp metering |
Sept. 29, 2022, 8:02 p.m. |
| Progress Report |
367_Progress_Report_2022-09-30 |
Sept. 29, 2022, 8:02 p.m. |
| Presentation |
TSMO 1 system simulation with ramp metering |
April 8, 2023, 10:59 a.m. |
| Progress Report |
367_Progress_Report_2023-03-31 |
April 8, 2023, 10:59 a.m. |
| Final Report |
367_-_Final_Report.pdf |
June 19, 2023, 12:02 p.m. |
Match Sources
No match sources!
Partners
| Name |
Type |
| Morgan State University |
Deployment Partner Deployment Partner |
| Maryland DOT |
Deployment & Equity Partner Deployment & Equity Partner |
