Tuesday, September 17, 2019

A Planner's Overview of Florida's Travel Demand Models

I've been applying travel demand models here in Florida for the last 22 years and been participating in the Model Task Force for at least a decade.  Florida is unique in that travel demand models are loosely managed at a statewide level.  We all use the same software (basically).  We have consistent naming conventions for filenames, processes, and variables.  We meet once or twice a year to discuss research that can go into improving the models.  We have a website that lists all of the different models that are available in each of the regions statewide.  Travel demand models are both an art and science, and you can tell who wrote or worked on each one from the internal code and overall structure--like telling a Monet from a Rembrandt.  Most of the models in the state are typical 4-step models with varying degrees of complexity in terms of transit choice and assignment routines.  Florida is really skilled at special generators--land uses that don't quite fit into the typical commuting pattern--because we have lots of them.  Some of the early data collection in the Orlando area used aerial photography to estimate vehicle counts at the theme parks.  I'm not sure we could even do that kind of data collection today both because the parks are so much larger and because of security and privacy issues.  At least 4 of the models in Florida are activity based models and they are painful to use.  I set one going for Tampa Bay (3029 zones) at 8 am this morning and it's not finished yet (at 6 pm)--and it's on a huge, fast computer.  In comparison, the CFRPM model (which is a 4-step model with about 5,000 zones) takes about 5 hours.  Going forward, we are working toward cloud computing for these projects, but security issues have slowed this considerably.  Model programmers are worth their weight in gold and the software isn't far behind in cost.  Cube cost me about $12,000 for the software about 4 years ago and they charge about 15% a year in maintenance.  Most of what I use it for is to generate trip distribution for projects that are in development review.  Nearly every jurisdiction requires a traffic study for new projects that mirrors the 4-step process (trip generation, distribution, (occasionally mode-split), and assignment) with operational analysis evaluated both for existing and projected conditions.      
Here are a few notes on applying travel demand models in real life planning:
1. Models are calibrated to roadway volumes based on the data inputs that are available.  That calibration is across the entire region, not based on any one roadway.  They are only supposed to be accurate enough to distinguish whether an additional lane is needed in the system.   Individual roadways may be off by 50% or more.  The bigger the roadway, the more likely it is to be accurate, but they can still be off by quite a bit.  Sometimes if the projection is off, we just look at the growth between the base year and the future year and apply that growth rate to existing counts to come up with the future base volume.  When we use that method, we also check that growth rate against the historic volumes on the roadway, with the caveat that we had national dip in traffic volumes between 2008 and about 2013 that affected every roadway, basically around the entire world.  
2. The farther out in time you look, the less trustworthy is the projection.  We generally create projections 15 to 20 years into the future, but take that with a grain of salt.  You can start thinking about a new corridor or roadway in that time frame, but you don't have enough information to take it to anything other than corridor location studies.  Five year projections are likely to be realistic across an entire corridor--it's probably ok to use those numbers for design traffic if you're careful.  If you're trying to project intersection counts or plan turn lane improvements, looking any farther than 2-4 years in advance is asking too much.  No one knows that much unless the area is very close to built-out.  I've done roadway projections 10 to 20 years in the future that I would trust, but only when I know everything that is going on around that location and I know that there is little that anyone can do outside of what is already built (and I whined the whole time doing it because it's just not a good thing to get people comfortable with). 
3. At some point it is important to manage expectations.  We may be able to do miracles with some help, but we don't do magic. Some things I can recalculate and make say what I want it to say.  Some things I can't.  If the volume projection is double or triple what the capacity is, then there's nothing short of new roadway that will make up the difference.  If it's within a few percentage points, people will adjust and there's usually tools to help make it work.  I have actually had a reviewer question a roadway projection because it was only 6 PM peak hour trips below the capacity.  He asked if I had made it do that on purpose and whether the number was realistic or the roadway needed improvement.  My response was, "Cool!  I didn't know it was that close."  I really didn't know it was that close--I had a check in the spreadsheet to see if it was over, but not how close it was.  The attorney for the client responded that if it were 6 trips over, they would make us pay for it, so forget charging us for it.  Such is the nature of development negotiations.
4.  Land use modelers and travel demand modelers are rarely the same group of people or even the same expertise.  Land use modelers are usually economic modelers that digest and process census trends.  Travel demand modelers are usually traffic geeks.  Yes, land use and transportation feed back on each other.  No, the interaction is too complex (to date) for those interactions to be understood, much less incorporated into the 4-step model.  Adding roadways to an area that is declining is like adding outflow pipes to an empty reservoir--the new pipe won't make any new water.  Adding a new pipe to an overflowing reservoir will make a huge difference both in the outflow and the reservoir (economic capacity).  Models can give you a good idea of what will happen in terms of traffic relocation, but it won't tell you how that will impact economic development unless things are growing so fast that congestion is getting in the way of growth.
5. Traffic forecasters are not psychic.  If there is a large redevelopment project that the local agency knows is in the works (has plans for and has approved) then it gets incorporated into the model.  Otherwise, redevelopment is rarely considered until the area is completely built out.  It's easier to build on empty land than rebuild on developed land. 
6. Growth in zones that are nearly built out may not actually occur because there's usually a reason it hasn't happened already--it is administratively contaminated and the local agency just hasn't realized it yet.  For instance, what happens when the required setbacks for a parcel are 30 feet from the property line and the property is only 70 feet wide?  It's really hard to build a 10 foot wide building and most developers won't even try an appeal.  This happens a lot in historic districts because any redevelopment or reconstruction means the property must be brought up to code, which may not be possible.  Unless the local agency identifies what is keeping orphan parcels from developing, it probably won't happen.
 7. Walking is a mode in the model but it doesn't count.  Same goes for biking.  Transit does get counted, but it's usually such a small percentage it doesn't even matter.  The rule of thumb we use in Florida is about 1.5% of all trips are transit trips, even in the areas with the best transit.  There's even a new Federal Transit Authority (FTA) process that reduces the complexity of the transit modeling because it's such a small impact.  Transit sections of the model are usually created to comply with FTA funding guidelines and analysis.  FTA models are even more complex than typical MPO models because they're asking for billions of dollars.  These models cost tens of millions instead of just millions.
8. Pass-by and diverted trips are not in the 4-step models in any large quantity, if at all.  Remember, the goal of most travel demand models is to predict roadway volumes.  Pass-by traffic doesn't impact roadway volumes.  
9.  Many of the simplifications that happen in the trip generation step are at least partially fixed in activity based models.  Activity based models also handle time of day better because it actually tracks travel by hour of day (or 15 minute period depending on the  model).  That's also why they take so long to run.  If you're running a model for a new project, it's not that hard to come up with the data to put in a new zone for that project in a 4-step model.  When it's an activity based model, things get strange.  Is your new subdivision going to have young families with small children?  Older families without children? Retirees?  You can model all of that in an activity based model, but making guesses for the percentages of people with different numbers of vehicles gets pretty  complicated. 
10.  Trip lengths depend on two things in the model:  average trip lengths for everyone in the model and the distance between origins and destinations.  It calibrates the outputs iteratively to get those two factors balanced.   As areas become more dense and congested, trip lengths (in time) will go down because things are closer together and go up because congestion increases the time to get places.  We hope those two effects balance out in the future.  Disruptive travel modes like TNC's and scooters are going to play havoc with our models in the future.  Our entire mode choice systems are going to have be completely revisited in the next 5 years based on the adoption of TNC's, bike-share, and scooter-share.  We haven't really even started to try to figure this out yet and it's a huge elephant sitting on the coffee table in the middle of the room. 
11.  There are calibration factors that can be used in the model to account for blighted areas, bad parts of town, rivers, bridges, or other cognitive/psychological factors that impact transportation, but modelers hate to use them because they don't trust them now or in the future.  River adjustments we trust, some days (but we don't like trusting them.)
12.  In Florida, the terminal time can be set as a factor for the zone or the area depending on the model.  Most models include the possibility to include long term and short term parking costs in the zone structure.  When you look through the socio-economic (SE) data, there are usually less that 5% of the zones that have a coded parking cost.  
13.  Intersection delay is only partially ignored.  The facility type designation often includes the number of signalized intersections per mile with a typical amount of delay per intersection, although it is not usually stated explicitly. 
14.  Model volumes can go way over capacity for individual segments.  The delay gets really bad, but it happens.  Most 4-step models have been daily models which means that an over-capacity segment just has more volume happening off-peak than normal (this is called peak-spreading).  Time of day models do this too, but not as much.  I assume (but don't know for sure) that activity based models can't be that flexible on their capacity designations.
15. Activity based models use a synthetic population.  They know the statistical characteristics of a zone and they create a set of households that conform to those statistics and then run them through activities.  It's not the real households, but it's close.  If I had a zone that was all kids and seniors, I'd worry, but that doesn't happen that often.
The Beimborn (1995) article has several suggestions that I thought were interesting:
1. Improved Bike/Ped Representation: California models currently have a much better representation of bicycle and pedestrian travel because they are required to.  Most of their models aggregate into zones, but include the parcel level data in the model along with all of the local streets.  That means they can actually assign the synthetic population to individual parcels and feed them out to the  zone centroid before they assign them to the roadway network if they're in a car.  Bike and ped trips generally stay within the zone itself, though in real life they can and do travel between zones.  This level of detail also helps better represent parcel access.  
2. They could use more trip purposes.  Maybe.  I have seen school trip purposes as a separate trip purpose in the Atlanta model and that would help some, particularly if it were coupled with the actual school district boundaries.  The CFRPM model actually changes the area type of the roadway (which changes the operating characteristics) based on the density/intensity of the SE data adjacent to that segment.  It's not the same, but it may have the same result as looking at different market segments.  Activity based models do market segmentation like the author describes automatically, but keep within the same 4 primary trip types.  
3. They could add land use feedback in the future.  (This is a "here, hold my beer" goal).  I'm not sure how well this can be done, if at all.  Not only is it a complicated technical issue, but also a subtle political issue.  Are modelers going to be comfortable projecting blight in an area (a decrease in future SE data) because a new facility has attracted that activity away from an existing developed area?  We know it happens.  No one wants to be the bearer of bad news.
4. Add intersection delays.  Can be done.   There are model platforms that do that now, particularly for small areas within larger regional models that are used for corridor studies.  Thankfully, we don't often do that here in Florida.  It's an issue of the accuracy you're trying to achieve with the overall model.  Remember, the model is only accurate across the whole model to +/- 1 lane.  That's a pretty wide margin of error.  Adding in intersection delays (that are already partially accounted for) would add another layer of complexity (read: computational time and expense) that would not add to the overall accuracy of the whole.  The technical term for this is "polishing the turd."  You're trying to take garbage and make it look like something better than it is.  
 I hope that helps.  It's a 50,000 foot view with a zoom lens.  If any of the modelers out there want to take issue with what I've said, I'm open to critique and will revise with the utmost of humility.  
Currans, K. (2017). Issues in trip generation methods for transportation impact estimation of land use development: a review and discussion of the state-of-the-art approaches. Journal of Planning Literature 32(4), 335-345.
Biemborn, Edward.  (1995).  A Transportation Modeling Primer. Milwaukee, WI: Center for Urban Transportation Studies.   Retrieved from http://www4.uwm.edu/cuts/primer.htm
Pihl, E. & Rousseau, G.  Introduction to Travel Demand Forecasting. Travel Model Improvement Portal (TMIP). Retrieved from https://tmip.org/content/introduction-travel-demand-forecasting

No comments:

Post a Comment