Document type sign assignment of partnership interest vermont fast
all right well good morning or good out and good afternoon everyone thanks for joining today's webinar on integrating transportation energy and emissions modeling across spatial and temporal scales my name is Mike Santino I'm the policy director for the National Center for sustainable transportation I'm going to be moderating the webinar today this is part of a monthly series that we put on and you should certainly stay tuned for upcoming webinars the National Center for sustainable transportation is part of the university transportation centers program which is administered by the US Department transportation we're a consortium on 60 universities around the country one of which is Georgia Tech and we focused on advancing environmentally sustainable transportation through cutting-edge research direct policy engagement and education of our future leaders we've got a mailing list so you can get notifications about our upcoming webinars and current research and other events that's at NC st dot UC Davis edu today's webinar will take an hour we have a presentation from Randy Gensler and then we're going to open it up to Q&A from all of you so think about your questions as we go along we have a Q&A feature in the Xoom webinar platform to type in your question you should be able to see it kind of on the bottom toolbar of your screen so feel free to type in your questions as the webinar or the presentation proceed we'll save our Q&A for the end but please get your question and questions you know if you think of them we will be recording this webinar so it'll be available along with the presentation on our website shortly after the webinar is over and with that I'd like to thank our welcome our guests and our speaker today Randy Gensler Randy's a professor of civil and environmental engineering at the Georgia Institute of Technology and Randy I will turn it over to you thank you thank you very much for inviting us to participate today I'm joined here by dr. hobbing Liu and most of the things that you'll see today he has made work everything when we talk about working with supercomputing cluster and Python scripts and the capabilities that we developed how bings been responsible for a reliance share of that also by digan Kim who is one of our senior PhD candidates he's been working with the air mod like expel dispersion model and you'll see some of the gridded applications that we develop later on so the presentation today really is going to focus on the integration of various models at different spatial scales and different temporal scales I'm using what we call noobs matrix which I'll talk about in a minute just fancy pre run version of moves through all the different possible inputs and a lot of that research was sponsored by the National Center for Sustainable Center for sustainable transportation or NC St but also a lot of it came out of an arpa-e project that we had with the Department of Energy about a three-year project number of million dollars to develop regional simulation modeling for the entire metro area and then to be able to link in energy and emissions module so we're going to talk about Luz matrix which we use for both energy and emissions in 49 states every state California I'll also mention that Georgia Tech fuel and emissions calculator which we use for some of the upstream emissions and energy activities really what it is it's the integration of both greet and miss matrix within a spreadsheet and online calculator and we'll talk about the applications across travel demand models traffic simulation models the micro simulation like pisum but also some the more macro simulation like dynamic traffic assignment and any kind of a simulation model that moves students around a fleet or individual vehicle can be connected with moves matrix data-driven real-time simulation and monitored corridor level activity and then finally we'll end up with some of these version models no mention a couple of the dissertations that have been recently published been doing there and you'll see agents coming out on let's talk about moves briefly a lot of you probably already are very familiar with moves moves is basically a modal emissions model in that you're actually calculating the vehicle specific power for the the vehicles that are operating on a roadway as a function of the each vehicle class and the characteristics of operation which include speed and acceleration but also it involves Road grade and most applications assume that the world is flat and so you'll see some some slides later on about we are moving into the realm of integrating great correctly and it does how it does make a different principle in terms of how some of those vehicles operate in other words the acceleration characteristics that some of these vehicles are different going uphill and downhill as you might suspect and that translates into differences in vehicles power and therefore energy use and emissions so the bottom line here is that because moves is translating your inputs into this VSP framework and then operating within the model and then translating these things back out into emission rates we wanted to develop a system that would allow us to do things a little bit easier than working with moves directly in each application moves models emissions and energy use across 13 different sources types and so when you run moves you're going to get different emission rates for classes and you'll get different emission rates for how those vehicles are operated if you look at the just a typical terms right in grams per second you can see that we've got pretty poor so 23 different operating bins across there they're numbered from 0 to 40 but there's some that are missing and really it's divided into a couple of different regimes you've got three different speed regimes a low medium and a high speed regime and then you have the VSP calculation using the equation up and as you move up in VSP you move along those bins from left to right you've got a braking bin you've got an idle bin and you got low speed posts and mediums and then as the VSP increases you can see that each of those curves upward to higher gram per second emission rate the the bottom line with moves is that if you can specify the way that the person sees what the vehicle fleet looks like and break those down into the individual vehicle classes and then you specify the model years for those vehicle classes and then you specify the way that those vehicles are being operated whether you have an average speed and a driving cycle that goes along with that or you have the individual second-by-second data you can actually calculate the VSP for those Roble operation as the federal chest procedure and what we have going on in the upper bound there is the speed second-by-second speed over time this vehicle is driving along on the laboratory treadmill and you can see that for every second of operation the VSP is being calculated and you have the bins of VSP operation down in the bottom so those correspond to the 23 different bins numbered from 0 to 40 and if you look on the very bottom line you can see the instantaneous speed and acceleration the current energy consumption total distance traveled and they trip energy accumulated over that time so we've basically been able to hook in with moves and those moves emission rates to be able to operate the characteristics the underlying characteristics that moves at different spatial and so if you've used moves whether you've used the old mobile model or any of the new moves models modeling you know complicated networks is really kind of a challenge if it becomes kind of tedious it requires that you generate different emission rates for different vehicle operating characteristics and traditionally modelers have either tried to simplify things by I don't want to say dumbing it down but by aggregation to say well you know we're going to assume that all vehicles are driving under a difference between these two driving cycles and we're not going to care about whether or not there's a difference between the different vehicle classes air operations on the road and maybe we will use one distribution of model years for all of the vehicles or at least for all of one's first type he'll either simplify the inputs or we create these look-up tables important that we do want to do and that generally consists of running moves for those scenarios let's say you know I want to have an emissions rate for 25 mile an hour average speeds on arterioles that apply in Cobb County because weed is different in Cobb County than elsewhere you pre run all of those different conditions that you want generate your lookup table and then apply the the lookup emission rates in the modeling that you're doing so the goal here with moves matrix really was to pre run move or all of the different combinations of input data so configure it for distributed computing so that we can take advantage of multiple cores on a computing cluster iterate the inputs across all of those different input combinations and run moves thousands of times for a region in advance of having to do the modeling and then compile all those outputs those energy use and emission rates into the multi dimensional matrix that's moves matrix and then finally develop a series of scripts and then apply these matrices to any analyses that we want to run for Metropolitan Area it takes about 30,000 moves runs to do the on-road exhaust emission rates and energy use so this is not the hot soak it's not the extended idle it's none of the other supplementals we'll talk a little bit more about that later on but it takes an awful lot of moves runs to do this it's across 21 calendar years in Atlanta we have a summer a winter fuel and a transition fuel for a number of months as we go from this number entering back we're doing 23 temperature bins so 5 degrees Fahrenheit increments and 21 humidity bins takes about 20 minutes per core per run so about 5 days to process the all of the different emission rates are all the different input files for metro Atlanta and that's given the but we get about eighty cores at a time that we're assigned by the computing cluster show you in a minute the bottom line is we're generating not hundreds and thousands and millions but five billion running emissions rates per region so it's a pretty good sized matrix it's a hundred twenty one gigabyte we've done more than a million lose modeling runs probably about 1.1 1.2 you what million modeling runs to date to generate the state of Georgia state of Texas Vermont a variety of different locations I should point out that this is assuming a fuel specific and program and assuming an inspection of maintenance program so for each modeling region they if they have a different fuel specification or inspection of maintenance program you have to do the 30,000 moves modeling runs it turns out that across the United States there's a hundred and seventeen different combinations of fuel specification and inspection of maintenance program though running lose matrix for the entire country is a little bit daunting on an 80 course before concluding cluster assignment so much larger computing clusters are needed to generate everything for the entire nation but we've been responding on demand as folks need applications we will develop the models for their their region and send them the results on a kernel hard drive a hundred twenty one gigabyte matrix and the process so the way that we do this is through what's called the pace supercomputing cluster it's a it's actually a co-op more than anything faculty members that want to participate in the coop they have a research project that will provide funding to purchase cores new course for the system I'm not sure if this is a sustainable system because as old cores go out new cores have to come in I'm not sure if this is a pyramid scheme or not but the bottom line is you buy your way into this and then you're assigned cores based on the number of cores that you've donated or contributed and then excess time on cores are made available to individuals as a function of being online and availability we have a new system that's just been implemented over at a new building that's been constructed on campus we're looking into how to get on to that system that's much larger and much faster but this is still an outstanding resource for us I would also say that one of the big benefits of this supercomputing cluster and I can't recommend it highly enough is the fact that they have staff that are that are stationed at Pace that work with the researchers so these are not as computing staff these are research staff that operate out of that system and we collaborate with them on a variety of different different projects so they're great to work with and provide a lot of really good support they also tell us which is another it's an advantage where the overall system not necessarily for us we have found that that moves is not structured as efficiently as it could be for operating on a super computing cluster it tends to consume more cores than they would like to have us consume and even though operating on multiple cores makes our individual runs go faster number of cores that are then not available for other folks to use is not in the best interest of the of the collaboration or the cooperative so we've been working with them to get down to a single core process or at least a two core process so that we can reduce the essentially the demand that we're putting on that system to do the runs it's run but overall the overall system will become more efficient by doing this they've been great to work so if you look at what comes out of that modeling process you do those 30,000 runs and you generate this 5 billion cell array of a mission race those emission rates are supported by calendar year fuel and I am program whole set and then the meteorology in that and before you do an operating run say you're trying to connect this matrix of emission rates with travel demand model or a simulation model you predefined what the calendar year is that you're going to look at or multiple years your fuel and inspection and maintenance program the temperature and the humidity and you pull it essentially a subset of that big huge matrix to work with and the Python scripts that folks here have developed do all of that once you have that and you have that subset it becomes a lot faster to work with and if you look on the bottom line of this chart the blue is the query that we're doing of those sub matrices so if you have the source type distributions you have the vehicle classes and then for each of those vehicle classes you have their model year distribution and for each of those model years you have the way that the vehicles are being operated either the operating mode bin distribution by VSP or an average speed and emissions rate for which we reassemble the moves you can go in and get the emission rates that you need so think of it this way but we as we really pre-run lose for all the different combinations that we would potentially be using nobody develops script go into that array and pull out the emission rates that we need and a link by link basis and the advantage of this is that we get exactly the same results as using noose where really all were doing is outputting all of the possible combinations and then we're reassembling those combinations in the exact same way that Louvre does before it gives you your grand per per second or grand per mile emissions rate and I don't know about you but I don't think I've ever worked with a project where the r-squared was 1 it is perfect which is great so the difference they're only associated with founding and that's not point zero one percent that point zero zero zero one we also developed for the Atlanta metro area not for any other regions what is known as moves matrix you point out we knew that we didn't have the start exhaust emission rates in their truck hood telling and evaporative emissions and we wanted them to be able to do regional inventory analysis for Atlanta for the metro area so yeah we have the running exhaust that we can do for anywhere but doing this for oh the start and soak and and and all of those emissions takes a lot longer as I recall it takes a
out 20 days and the supercomputing cluster to generate the matrices for a region and the matrices are huge they're absolutely huge but we have done it and if we're capable of doing it for any region it's just a matter of having enough supercomputing time in order to generate the rates or for other folks and again get the r-squared of one it gives you a lot of flexibility for use in scenario analysis though when you do the you have this inventory to work with so you know for example if I want to know what the on-road emissions are that's a lot easier but if I want to look at a new plan a transportation plan for the metro area with a build-out for 20 40 and they're they're talking about making major modifications in the structure of the transportation system we can do the on-road fairly easily but now we can also bring in the exhaust and the evaporative emissions and do the poem which is what the regional planners QP so before we move to some of the applications I do want to mention that we've also developed what we call the fuel and emissions calculator this is an NC ST sponsored project it originally started with a funded project we did for the Federal Transit Administration and so it's a culmination of moves moves matrix did you rate the on-road emissions right that's the on-road pumped a wheel but also integrates Bri originally only for those transit vehicles so that you could do all the upstream well to pump energy use of emissions so you could look at you know what happens if you should have part of your fleet new CNG the on-road changes but also the upstream energy use and emissions change so by integrating lose matrix where you do get the ability to put in a driving cycle and have the VSP be calculated for your fleet and then also integrate the upstream gives you some advantages in various analyses you might want to do with respect to alternative fuels this it like I said it was originally for does include capital costs operating and maintenance costs energy use and emissions does include hybrids and electric vehicles so some of the things that are missing with moves you can generate all the upstream here and we put in estimators into the fuel and emissions calculator for those vehicles and this should be ready this next semester to be updated based on some of the dissertation research that I'll talk about at the very end of the page but it's easy to use it's available online you can download it play around with it send us any comments that you have and the nice thing about this calculator is that you can see every single equation that's being used it's all open source so there's there's nothing going on but you know behind the firewall or in a black box you can see everything that the model is doing and you can wait if you're one of these different driving cycle you want to use different assumptions about F dream energies then you're capable of doing them so with that let's turn to some of the applications that we've developed you've got this this on-road emissions matrix and you've got the the full regional emissions inventory matrix available and we've applied these at every spatial and temporal ski we can think the models were operating here in the metro area I'll talk about the regional travel the man model Atlanta regional Commission's activity based model which is actually one of the probably the top five models in terms of complexity in the United States we've linked with that we do a lot of corridor scenario analysis we've done some of that for the state road and totally Authority associated with a conversion of a carpool lane to a managed toll lane with an increase in vehicle occupancy we connect directly with nism and some other micro state microscopic simulation models our preference is this and primarily because it's got really good visualization tools but it's capable of being connected with any simulation mode you want we've done and then outputs as well and then we'll talk about microscopy pollutant dispersion modeling connections we've done caroline 3 we've done right now and by the time dajin is done we'll have connected with all five of the dispersion models including our line no no little app component for the RPE project which allowed us to look at energy and emissions use in real time and pre-planning of travel across different routes and time of departure and end mode so including even park-and-ride to transit well you can look at I mentioned also the fuel emissions calculator and the cost calculator can be we have a few different models there that I didn't mention the cost calculator because I take the flight out for time but we have a really good online cost calculator that is available through n CST as well so you can see how much it costs to own and operate your vehicle over the life of the vehicle instead from lab anyway these other two models can be applied in series with mousse matrix so it gives you a lot of powerful capabilities in terms of looking at changes in energy use operating cost and the like okay so moves matrix 2.0 is what connects to the travel demand model we modeled the 20 plus County metro area some of these counties are got a couple new counties that are coming in the main planning area the biggest population densities are in the white there and the newer counties that have been added over time as the metro area has expanded are the counties and brown over to the right the model is pretty complicated there it's mostly choice based or all the algorithms so even the number of trips generated per household is a choice the probability that you'll make one trips two trips etc the trip distribution model is okay so it's the probability you'll go to a zone over Tibet then or if these purses are sewn to the southwest so not just a straight gravity model traffic assignment is pretty much the same as as we've had in the past using Frank wolf algorithms to distribute those trips to the network but then no choice based model so it's quite complicated the model that is currently in circulation has 5800 zones transportation analysis zones and about 74,000 network weights in the Lincoln node structure the new model that was coming out the one that we've been working with for most of our applications as 202,000 links fortunately we learned the same thing that AARC has learned and that is 200,000 links is too many links the latest model that will be coming out will have about a hundred and thirty-eight thousand links in the next generation but also the number of analysis zones are increasing so for any of you who run a travel to male model you can recognize these are very large numbers this is a model that runs overnight more than 24 hours to do the modeling runs so scenario analysis is a little bit cumbersome but the nice thing is that by being able to connect directly to moves matrix again moves matrix 2.0 you can then take the outputs of that model and automatically generate the emissions inventories and you can do it by by pollutant type and you can do it also by source type within the pollutant type so we've got the output for the on network and off network emissions for the VMT hydrocarbons and PM and you can see that you've got the start emissions in the lower left and then the parked vehicle evaporative emissions particles in the evaporative emissions for hydrocarbons PM 2.5 for the off network so it allows you to then look at where these emissions are occurring so you got your spatial component and your temporal component because the travel amount model is actually saving every single trip 20-plus million trips in the metro area across the day and then by running that model and applying the lookups and moves matrix you can break this down by hydrocarbon source type so start exhaust have happened to permeation of the gas lines fuel leaks cetera and on the PM side over there you got brake wear tire we're running exhaust all of these can be automatically generated by linking these two models together our ultimate goal really is to be able to take any scenario analysis that comes out and automatically run it with with moves matrix in background so that the next morning when the modelers come back in they have not only the outputs from the travel demand model but we've automatically generated all of the energy use and emissions that go along with so microscopic simulation modeling is the next visum is one of those more complicated models in a varves involves car following algorithms so each vehicle is moving through the system this one actually tracks the location and the position and the time of that vehicles movements and it's integrating things like weaving and and distance separation braking and all of that and so if you integrate through the combi interface which mike hunter and his crew upstairs do a lot of that you can basically retrieve the vehicle trace data out of the network file so you can see for every single vehicle with the speed and acceleration characteristics are of that vehicle as it's moving through the system and so that it gives you a lot of capabilities with clearly directly calculating the VSP and STP for the heavy duty vehicles and then applying the the energy and emission rates from moves matrix directly now there is this M doesn't assign directly the same source types that moves uses you get based on your inputs you get the light duty vehicles and so there is an overlay that we do in order to match up they flee characteristics with what we're getting out of the commenter face we do have a little interface that we use in there to basically break it down by model years and and vehicle classes for moons modeling but it's a pretty straightforward process that can be modified then we process each of those second-by-second traces to BSP imagine the individual vehicles with the energy and emission rates that go with it and then again through the communist we append their data back to the to the trace data so you get in the output file you then have the ability to look at the animation of energy consumption or emissions this is for co2 emissions so energy consumption as well you can see that we've got the individual vehicles being assigned to the network you've got the motorcycle as the diamonds light duty vehicle so there's circles and you can see the combination trucks of the Box this and then over on the right we've got the coating that goes with the color coating that goes with the emission rates and grams per second and if you look closely at this you can see that you have emissions increasing during the acceleration phase you have emissions increasing at high speeds but really what we're doing is calculating the second-by-second vsp for all of those vehicles applying the emissions rate and then animating that output so you can use it both for comparative analysis for scenarios like you're going to change the highway operations design you're gonna add a lane subtract the lane change a weaving section all of those things that this and will allow you to look at and then you can generate the energy and emissions changes that go along air mod is a is the pollutant dispersion model that we're currently using this is what Dai Jin has been working with on the supercomputing cluster we use it both for air quality impact assessment for things like environmental impact assessment comparing across alternatives for conformity analysis and for helping back assessment so the goal here really because air mod is another one of those models it takes a while to run is to try to speed up the operating characteristics of of the predictions of downwind pollutant dispersion from vehicle activity and then to also do some link screening which will come back around to basically ignore those links that aren't contributing emissions at a receptor that you're interested in no air mod for micro scale dispersion is available free for anybody to use and we'll talk about about we do the worst case pollutant concentrations and identify huh so what's going on here is a linkage between the traffic volumes observed on Jimmy Carter Boulevard and the hourly Co concentrations that resolves for a weekday in 2012 where the background concentrations are excluded and the concentrations are not high they're not super high they get up quite a bit up at the at the peak of the peak and in terms of the regional scale applications that we've been doing and we'll talk about how we speed up those processes in a minute if you look at the time it takes to essentially look at the impact of every single link on every single receptor where you've got a grid of receptors over the metro area it takes about ten days to process on the super cluster so we got an awful lot of links that we're talking about the breakdown to 160 1000 individual polygons for air mod analysis you got a million receptors that we overlaid on this grid so it's quite a computational process you could get a worst-case analysis using the highest emissions rate and selected and it only takes about three days to process that but you have the capability of both doing a regional analysis over a long period or doing a short-term analysis to get those pollutant concentrations before we move on to the next step I do want to mention that one of the things that we've been doing is working with grain and part of the helping Lou's dissertation work a couple years back was to develop a process and procedure for generating Road grade at very high resolution over the transportation network in Georgia we've got for Atlanta alone we have about fourteen hundred miles of freeways 7,500 miles of major arterials and 12,000 miles of local roads and using the US geological surveys digital elevation and a variety of techniques for smoothing both for filtering and smoothing and inserting roadways where you have bridges over culverts and things like that that process allows us to get about a half a percent grade accuracy on the arterioles and down to around 0.25 percent on the freeway so this is really good quality grade data at very high resolution on those freeways and the arterioles so having the availability of this high-resolution digital elevation model you can take any second-by-second data from observations in the Atlanta metro area and you can assign for every second of operation not just the speed and acceleration that you may have monitored or simulated but now you can assign the grade to that place on the actual roadway grade for that location and then stood up at the very beginning that it does matter with respect to speed and acceleration characteristics what the operator of what's art what the grade is so in other words when your scuse me when you're looking at heavy-duty trucks for example the upgrade is different than the downgrade I'm not as much for the light duty vehicles they operate pretty pretty similarly knowing what the grade is in terms of the VSP calculations does matter across the board here knowing the difference between the speed and accelerations across the grade distribution say uphill versus downhill not nearly as important for the light duty vehicles but does become important for the heavy duty so this is one of the applications that we've developed with the activity data that was collected on a second-by-second basis for the metro area as part of the atlanta regional commissions travel diary surveys that they did in 2011 so we actually had about 500 vehicles that were monitored second-by-second and then we collected data from a variety of Marta bus lines and different activity characteristics on different routes so it does make a difference and you'll see that if you look at the energy consumption rate the looking at the true emissions which is in the blue there versus assuming that grade is at 0% and then assuming that the activity is independent of the uphill versus the downhill you can see that it does make a difference we're talking about significant percentage difference there and so we recommend the use of grade first off we recommend the use of grade everywhere in in our analyses both for energy modeling emissions modeling and certainly for the microscope dispersion modeling you can see on the right that we've got differences emissions as a function of concentrations that result as a function about Billund downhill grade and then like I said we certainly reco
mend it for air mod for those areas where you have some significant grades it does make a difference and you can see the differences here's different same location but two different runs and the upper left you've got without Road grades everything's flat on the bottom left you've got the with Road grades you have the percentage of Road grade then you have your predictions in terms of PM 2.5 concentrations of air mod that Center upper one is for flat grade and the lower is when you include the grades but there are some significant differences result around three micrograms per cubic meter or so is to see the last thing I want to mention with respect to gridded modeling because we again when you think about it if if you can do all of your air mod analyses for a transportation plan at once right if you can look at every single location in the metro area by running everything in background on a super computing cluster so that when the Atlanta Regional Commission finishes running their scenario they get the emissions inventory for the region broken down by all those different characteristics and then they can also get the predicted pollutant concentrations across the entire region it makes our planning processes much much more agile so here we have a comparison of what you can't when you have a standard 200 by 200 meter grid which is up in the upper left there the applications have been done with gritted emissions modeling and concentration modeling use these kind of sparse grids you get the pollutant concentrations that you note there in the center of if you go in and you define a very very refined gridded Network where that gridded network is is kind of intelligently laid out it's laid out with respect to having more receptors closer to the roadway diminishing the need for those receptors as you move away from the roadway because the concentration profile drops down significantly you end up with the concentrations in the in the center bottom box and that's a big difference so running things that are very refined scale is what we also recommend but the problem with doing that is that you take an awful long time to run on the super couponing cluster when you're looking at every single sceptre that we've got in the lower-left and every single link that's involved in there so one of the things that I jen has been working with is called supervised link screening and this is a tool that is essentially trying to do one thing and that is to eliminate link receptor pairs where there's just not a significant impact from that link on that receptor now a long time ago we did it for cow line three and for the old moves model by doing a mass flux comparison it was okay with that but with machine learning and you can really develop some refined screening tools the bottom line here for what dodging has done so far is that through that screening we can reduce the air mod pace so getting run times by almost 99 so it's a big difference in terms of being able to run those models quickly and you're still getting 95 to 97% of the concentration by being explained by that slimmed-down model compared to the oil ink model so with appropriate interpretation of those outputs we think this is the way to go for doing essentially microscopic or micro scale modelling over the entire metro area and here's your your fits for both downtown Atlanta and northwest Atlanta and these are fits that given the the accuracy and and in that of the models I can live with these these are really good results so before we wrap up here I'd like to talk a little bit more about some of the ongoing modeling work we have going we've got two new dissertations that are out one is on the use of of autonomy to essentially generate VSP compatible emission rates for hybrid electric vehicles so that we can incorporate the modal based activity for plug-in hybrids hybrid electric vehicles by different type of hybrid and battery electric vehicles into the the moves framework and then the second one is it uses the same kind of approach for the transit vehicles but also includes some optimization models forthcoming we have some distributive justice assessment tools coming for planning where we look at pollutant concentrations impact on folks in the metro area and and looking at at the outputs of the model with respect to whose activity is occurring and who's being exposed by income household size and all the other demographic because news doesn't currently include battery electric vehicles on the line we wanted to have some kind of a model that would be comparable to all of that and so should add as you did her modeling on this and essentially what we're doing is predicting the modal energy use for those on-road operating conditions I wanted it to be reasonably accurate which it is and in fairly parsimonious we don't have a ton of different variables included in here though the conceptual framework is to try to figure out alright we've got an energy source and energy conversion we got a combustion engine or electric motor a combination of those two and using the autonomy model which is a full vehicle simulation model what Jaudon did was to essentially develop a Bayesian series of models to predict what the operating mode of that vehicle is whether the engine is on or off whether the electric motor is being batteries are being recharged or if the batteries are trading just basically create a modeling tool to predict the operating mode and then for those operating modes using autonomy outputs for 50,000 second-by-second trips they develop essentially a VST binning machine learning process to develop ESP binned emissions rates and energy use across both fuel consumption hybrids as well as electricity consumption and have to refer you to the full dissertation to get the entire body of work it's a but the bottom line is that it does a really decent job for the fuel consumption we can do better on the electricity consumption but really that comes down to the specification of the vehicles that we've used in the autonomy modeling so yeah I know we are predicting predictions right we're taking a simulation model and we're taking those outputs and then we're predicting as efficiently as we can those simulation model outputs so that we can use them in that in the moves bsp bidding approach but we think it's much better than what we have right now which is much nothing this is really good and it's gonna be excuse me some enhancements of this coming down the road in looking at the application kind of scalability components of this she applied this again to the metro area she used the travel demand model outputs in terms of the trips and the origin destinations and then used an interface with with one of the Google tools to essentially assign routes and then looked at the actual energy consumption being predicted and the bottom line is we're going to get a pretty decent reduction in in energy use with the electrification of the fleet and it was being battery electric or any kind of any electric vehicle using the assumed a mix into the future we get about a 30% reduction in energy used so stay tuned we'll be online the last tool is is what we call roadway sim and transit sim this is developed for the arpa-e project we essentially took the 200,000 link roadway Network code read it appropriately in Python modeling to do Dijkstra's shortest path algorithms integrated the 90 plus Marta bus and rail rates and the 23 Greta express bus routes using GTFS data for routes and departure times and for station locations and with this process basically we developed a system where you can pick any origin pick any destination and the system will give you the shortest path from origin to destination by any mode you choose so you can drive direct you could carpool to or you could take your vehicle to a transit station do Park & Ride and get on transit but it allows you to be sure to do the shortest path modeling through the network and so that modeling capability was translated into this new framework developed by n Lee for optimizing energy consumption with alternative fueled transit vehicle transit fleets and so the model is now capable of predicting transit vehicle energy use again using the moves matrix components and upstream greet components the simplified version of the of the alternative fuel or the electric vehicle and simulations is really good r-squared is close to 1 and the mean absolute percentage error is less than 3% and it allows you then to assess exit vehicle performance for an array of fuels that gets using like I'm gonna buy a hybrid bus or electric and the routes that they're being assigned to one of the problems that we have in Atlanta is that our we have a lot of great and some of our transit routes are quite long and battery electric vehicles are simply not compatible with all routes here but this modeling tool allows you to look at the assignment of the buses for the routes the impact of station placement a recharging station placement the impact of changing routes to make them more efficient assigning buses to specific people it's really a fantastic model a lot of technical detail and a lot of different options to work so again in the upper box we're talking about the use of metonymy simulation to generate the micro trip energy consumption rates and then using the machine learning to come up with that final model for applications but it's much more efficient process and allows you to do a variety of scenarios and so the scenarios that we have the simplest term is just basically looking at what happens if I purchase an electric bus or hyper and assign it to different routes which is the best in terms of reduced energy use you can then optimize bus route and Depot assignment so I think of it as I'm going to purchase 40 vehicles and I want to know where should those buses be assigned based on a most efficient place into charging stations which automatically figure out for you as well and route those vehicles it's fine and the last one which I was really pleased that she was able to do because it was so hard was to optimize the annual fleet purchase and assignment so now if you have a budget you have let's say you have 1.2 million dollars for the dollars for procurement of new vehicles it will tell you what vehicles to purchase and where to put those charging stations and if you have extra leftover money it may tell you to go ahead and buy that charging station that you're going to use next year so it's another one that I really encourage folks to read so I guess in summary news matrix is it's not rocket science it's brute force moves modeling we are just pre running that model thirty-eight thousand times and creating five billion emission rates to look up so you get the exact same energy use and emission rates it is a pre run process that allows you then to work much more efficiently with the output the processing time to generate an emissions rate is 200 times fast if you don't have to do any of the coding ahead of time to figure out what your your inputs are going to be for the moves model you just have to have the inputs for pulling the the results so it's a lot easier a lot faster it could be applied to any spatial and temporal scale we can quibble about whether you should apply it to individual vehicles because it wasn't really designed for that but that's fine if you want to look at comparisons of what the vehicles are doing on a corridor as a function of changes we think that it's still appropriate new regional analysis corridor case studies simulation modeling integrate two apps it's really quite flexible the scripts will then link all these traditional models with the simulation models we've got instructions for how to operate those scripts and how to develop your own and finally we are linking this with dispersion model so especially airmid grid right now I thought cryptic decoding is required we couldn't be doing this if we didn't have folks in the lab that can program in Python in terms of operating with the models expertise in distributed computing GIS GIS visualization tools and all that helps but you know we've worked with a number of customers we've got a consulting firm that's been using the model we're working now with Texas A&M I know that Vermont has used the model Iowa State so you know we're here to help with respect to the implementation and all we ask is that you send us a hard time for us to for you and I guess the last thing is that we've got you know these new energy optimization models and that that we're developing based on the system's that we've created lots of new deep learning applications one of our PhD students Anya Lu is not with us right here today because he's sitting over in his final exam review for machine learning so there's a lot of machine learning that's going on in the lab so with that I think I'd like to conclude and we could take any questions that folks have great thanks a lot for that presentation in these long spine packs so once again folks if you have a question to ask we ask that you use that Q&A feature that you should find on the on the bottom of your screen on that toolbar and just type it in there so we've got a few questions coming in randy maybe first of all we'll start with the simplest could you just clarify in terms of outputs of loose matrix which the types of emissions the outputs is that is it greenhouse gases is that I think you had a slide about particulate matter NOx well what all do you get out of it all of them everything that moves produces we take as an output that's why the matrices are so large you're getting all kinds of stuff that most people will never use but yeah you get all of the different pollutants that moves is capable about great Thanks we got a couple of questions about how lose matrix factors in congestion and whether it can account for Geographic or temporal variations and level of congestion okay so think of moves matrix is just the operates from moves right so with respect to congestion your congestion shows up either as a change in average speed or a roadway facility type and let's say that I was looking at two different scenarios one where it's congested right now but I'm gonna do some kind of a modification to relieve congestion I may go from an average speed of thirty six miles an hour on the freeway according to the travel man model or your simulation model to an increase to say 45 miles per hour on the freeway well moves itself if you specify the average speed and facility type what moves is actually doing is it's picking two driving cycles from its library that found that so let's say 36 miles an hour is the current condition and maybe looking at an internal speed and acceleration profile a driving cycle for 30 miles an hour and another one for forty two miles an hour and that's weighting those two together and moves is actually generating the vspeed in distribution for those two combining them and weighting them and then you get the emissions rate coming out for your 45 mile an hour or 36 mile an hour so using average speed and facility type you get the same results that moves would predict for congestion if you have observational data let's say you have speed and acceleration profiles from monitored vehicles or radar guns or what have you you could actually generate those those VSP bins yourself and you can do that analysis but either way it's using the input that you give it right it's just taking the results and saying okay I need the emission rates for these conditions and then goes each then go and get what you need from this matrix so lose matrix doesn't do anything itself it's the interactive scripts between your model and moves matrix they give you those outputs but yes it does incorporate congestion that way that makes anything so we also got a question about let's see since it says they have the city or an urban area was contemplating a scenario with vastly fewer light duty vehicles transitioning to other types of transportation the question is whether moves matrix would be and having you to kind of compare the energy and emissions profile of a scenario like that where maybe you know fewer light-duty views you know passenger vehicles and more because I think I think I saw in an early slide you also have the profiles for l
ke transit buses and intercity buses so so if you were comparing scenarios where you were kind of on and on a scale of an urban area you were you were changing or adjusting the the use of those types of modes would would move matrix B allow you to kind of compare the the emissions profiles of those beautiful areas yeah it would be add a few cautions here so you can look at it from the perspective of the travel of man model and the travel of man model is not terribly accurate at predicting what's going to happen on an individual roadway link there's a lot of folks who basically say well we're gonna run the travel of man model and then we're also gonna run dynamic traffic assignment while we're running that model either with the outputs or as an integrated process in order to get a better estimate of how those vehicles are going to distribute on the network and what the speed and acceleration characteristics of those vehicles are going to be so yes you can do it with the travel man model but you know your your outputs for being able to assess what's going to happen at a microscale level is not great you're really better moving to dynamic traffic assignment if you have faith that that is the you know a good solid model you must calibrate it for your region you can't just take something off the shelf and run it because it involves signal timing and all kinds of other assumptions that go in there they're really hard to keep calibrated or better yet if you're looking at a smaller area you can go to something like this and you can look at the traffic stimulation that will result give you an example we have 10th street in Atlanta where we've taken one of the roadway links away and created a protected bike lane and there's been a change I mean you have people there using the protected bike lane but you also have a change in the operating characteristics of the vehicles now on that roadway some of them have diverted to other routes which may have different driving characteristics and those that have stayed on that route lower travel time and more stop-and-go activity you can look at those trade-offs with respect to the system but it's really developing the good estimate of what those traffic changes are that's e to all of this you can have a change in traffic volumes you can have a change in on-road operations you can have a shift of travel to transit in which case you know things will generally be more energy efficient one if you have a good ridership on those transit vehicles but all of those things really really center around the analytical work that you do on the changes and vehicle activity moves again is your tool for linking those two things together so you have to have good activity change results to go with that now moves does not include you know bicycles doesn't include electric scooters or you know any of these other vehicles but we've done off model tools for the RBE project we're really straightforward for looking at energy consumption of it's not terribly difficult to do that if you really want to but again change it at the scenario analysis most important feature removes matrix is just your tool to get the appropriate admission rates to go with those great ok so the next question and I'm just gonna read verbatim because it might be a little out of my desk how do you estimate the distribution of vehicle power trains ages engine displacement and other ESP characteristics yes so the first first element is really the the source types the source side distribution now what we have done and you have a variety of different approaches you can do for this you can look at what the travel demand model and and air quality planning agencies are currently using for the fleet and that's exactly what we did when we did the scenario analysis for the region we wanted to use the same fleet that they use that said the fleet is not I mean you can't really apply an average fleet across the region and get reasonable results I give you an example when we commute from the suburbs I don't I live in Midtown but if you commute from the suburbs to downtown you don't take your average vehicle you take your new vehicle your newest most reliable vehicle most comfortable and it tends to be a higher model year that tends to be a newer newer car on the other hand if you're looking at traffic on arterioles and on weekends used tend to see a lot more of the older model years though we're big fans of doing license plate data collection and working with the agencies that have access to those data to get the make model involve your distributions or the fleets that are on the roadway if you have those they match directly with what you can do and move the source type the model years can come from those license plate studies otherwise you're generally using the registration mix weighted by the annual mileage accumulation rate now and with respect to all Road operations if you're working at the higher level where you're just saying I want the average speed and facility type we've talked about that you just specify that and you you get the BSB weighted for those operations you know how those vehicles are operated let's say you're out there with laser guns collecting data on on freeways or your car doing car following or you're doing instrumentation of those vehicles you can take those speed and acceleration characteristics and use them as well so think it moves that's really your thing what is my fleet you know that cars versus trucks versus different kinds of heavy duty trucks and what's the model year distribution within each of those depending on what data sources you have you can specify those combinations and then finally what is the on-road operating characteristic either the second-by-second driving traces that the vehicles are following or a vspeed and distribution or just average speed and facility type those are the three components that you need to be able to pair up together and again there's a variety of different resources that you can use these great well that takes us to the end of our time Randy I'd like to thank you for giving the great presentation and thank all of our participants for joining us and for your great questions please look for a follow-up email that contains the quick survey be really great to get your thoughts on what we can do better for next time and also asking for some feedback on topics you'd like to see featured in future webinars so we'd appreciate if you take a couple minutes to fill that out and with that we'd like to thank you and wish you all a good rest of your day thank you
