1372284 - Impact of On-line Traffic Signal Optimization on Operation and Energy Performance

In the second phase of Regional Mobility project funded by US DOE at Chattanooga, TN, the project team was trying to control a scaled-up area than an arterial from the first phase of the project. The area to be controlled by the team from the National Renewable Energy Laboratory (NREL) covered 39 signalized intersections. We selected the online control algorithm, Virtual-Phase-Link-based Model Predictive Control (VPL-based MPC), that was used to control six signalized intersections in the first phase. The scaled-up control experiment included more scenarios (e.g., five-leg intersections, two intersections controlled by one controller with overlaps, less observability, etc.).

In this work, we developed a scaled-up simulation in SUMO to test the control algorithm. We tested both morning peak demand and afternoon peak demand. The base SUMO model was developed in three stages: network development, existing signal timing plan implementation, and traffic assignment. City of Chattanooga provided the base regional network and configuration files exported from the field controllers. The study area network was extracted from the regional network, and the network discrepancies with the real-world network were verified and corrected through checking Google Earth satellite maps. Traffic Control Interface (TraCI) in SUMO was used to implement existing signal timing plans in SUMO. We leveraged the newly developed NEMA Type controllers in SUMO that was developed by NREL in the simulation. We used estimated volume with TomTom data as the ground truth volume and calibrated the Origin-Destination (OD) matrix as the demand for SUMO via a Genetic Algorithm (GA) running on NREL’s High Performance Computing system.

The field traffic signal control used a time-of-day (TOD) plan to address the variation in traffic demands. This study implemented the VPL-based MPC to optimize green splits during the simulation. This method used the number of vehicles in each virtual phase-link as input state variables. The input state can be extracted directly from the simulation. However, most intersections in the field does not have sensing technologies that can provide estimate of such information. We purchased real-time probe vehicle data to estimate the input state variables. Therefore, we also tested a proof of concept method using simulated probe vehicle information to test the impact of the imperfect state estimation from probe data.

We evaluated the performance of the control in terms of mobility and energy consumptions. Vehicle trajectory and link-wise signal performance measures were collected from SUMO for the evaluations. Simulation results suggest that the VPL-based MPC with ground truth state inputs experienced lower total network-wide delay than TOD plan by 35% in the morning and 30% in the evening peak. The VPL-based MPC with estimated the number of vehicles from probe vehicles experienced 2% more total delays in the morning peak and 16% less total delay than TOD in the evening peak. Mean speed and mean fuel consumption results also suggest the performance benefits of MPC methods over TOD while imperfect state estimation may hurt the performance.