Overview

Concurrent to the TRB IDEA implementations in Seattle and Atlanta, the same passive data synthesis process was implemented in the smaller region of Asheville, North Carolina. This was done so that the synthesis process could be externally validated against volumes as is done traditionally during calibration/validation. Additionally, two planning applications were implemented using the passive data model and MATSim combination: a bridge closure and a tolling scenario. This comparison could be done in a more timely manner in Asheville than it could be done in Seattle. The next application will include modeling several automated vehicle scenarios with future population forecasts.

The following flowchart shows diagrammatically how the synthesis process, which is referred to as a discrete-event simulation (DES) here, compares to the four-step model that Asheville metropolitan region currently uses. For context, an activity-based model (ABM) process is shown too. Both the four-step and activity-based models essentially use the same type of inputs. In the case of ABMs, the inputs are more detailed. Although ABMs estimate more sophisticated behavioral dynamics than trip-based models, they both flow through sequential steps, where matrices of information are passed between the steps. In either case, the demand model outputs get summed up and feed into an assignment model, which can be either a static-assignment or a dynamic one.1 The assignment model loops back to the demand model until equilibrium is met (not shown in the flowchart).

In the case of this DES approach, a different set of inputs are used. The inputs rely on passively-collected “big” data. In the demand model, person and firm populations are synthesized with consumer and firm data as the seeds and, in a similar way to population synthesis for ABMs, is controlled with U.S. Census data totals. Then rather than sequentially estimating large matrices for each step, a DES is used. DESs are a way to model the behavior of a complex system as ordered events along a timeline, where each event occurs at a particular instant in time and marks a change of state in the system. For example, a change in the state of the system might be leaving one’s house or arriving at one’s work. The DES is trained using tour patterns from NHTS, and in the future will be trained with other passive data to reflect up-to-date and local changes in tour behavior. In other words, the DES models the same aspects of behavior you find in four-step or ABMs, but with a unified simulation rather than sequential steps.2 It then passes the demand trip tables into an assignment model.

An example of the trip table that passes from the DES to the assignment model is show below with one random individual’s travel diary from Asheville. This table is relatable to person and household sociodemographic data via per_id and to firm data via place_id. The start and end columns are shown in military time.


In Asheville, the DES was fed first into the existing static-assignment model used in the region, which was built in TransCAD. Following that, the DES was fed into MATSim, an agent-based assignment model (similar to a discrete event simulation). No feedback loops were used in either case, although MATSim does have a replanning phase within the model itself that maximizes utility for each individuals’ schedule. In the future, one could imagine a single agent-based model that would do both demand and assignment in “planning” and “real-time” layers.

Validation

The passive data demand model with both the static-assignment and agent-based assignment models were compared against traffic counts. Alongside, the existing regional four-step model with the static-assignment model are shown. Note that the passive data model currently only handles internal trips to date; demand tables for commercial vehicles, visitors, and external trips were carried over directly from the four-step model.

Reminder
The existing four-step/static-assignment model is calibrated specifically to the local traffic counts whereas the two passive data implementations are not calibrated to Asheville.

Assignment Results

The passive data model produced effectively equivalent levels of accuracy when compared with the current four-step/static model. The following scatterplot and table summarize the results. In the left column of the scatterplot, the reference four-step model validation is depicted. In the right two columns, the DES models are depicted. The top row illustrates the count scatterplots, and the bottom row illustrates the maximum desirable deviation tolerances. The points are color coded by facility type. Note that 41 and 47 links in the static-assignment models, respectively had percent error greater than 200%. However, these were all on Other Major or Minor Thoroughfares with AWDT volumes under 3,000 (shown with a vertical dashed line). For visual clarity, they are not shown.

The largest deviation from the maximum desirable deviation tolerances are with expressways in the DES+MATSim model. In this case, the resident data only are fed into MATSim. When agents figure out their plans, they do not see any other traffic from commercial, external, or visitor trips. They will fill up the expressways first. Post MATSim, the commercial, external, and visitor trips from the TransCAD assignment are added to the MATSim network results. Because of this simplification, the expressway assignment is expected to be high.

Planning Applications

In the following planning applications in Asheville, the DES and MATSim are used.

Bridge Closure

In this application, a bridge link on Amboy Road was deleted in both directions from the network. The following map shows where the bridge is in relation to downtown Asheville.


The following network difference plot shows the effect of the bridge closure. Red/yellow mean fewer trips (where full red means a loss of almost all of the links’ volume), blue means more, and grey means unchanged. The difference is relative to the base scenario with the bridge open (i.e., \((V_{no bridge} - V_{base}) / V_{base}\)).


Tolling

In this tolling scenario, a toll of 10 cents per mile from 7am to 9am and 4pm to 6pm was implemented along the I-40/I-240 loop, which can be seen in the map above at the default zoom level. The following relative network difference plots show the network just before the morning toll begins (around 6:40am) and just after (around 7:00am). Again, yellow indicates a decrease in volume whereas red would indicate a change to zero volume.




  1. ARC is working on integrating their ABM with a dynamic traffic assignment model.

  2. Mode choice is being tested in both the DES and in MATSim.