TRAFIK I TÄTA MILJÖER Svante Berglund (Modellutveckling, text och scenario) Patryk Larek (Modellutveckling, text och scenario) Joel Franklin (Scenario)

(1)

WSP Analys & Strategi

121 88 Stockholm-Globen Besök: Arenavägen 7 Tel: +46 10 7225000 WSP Sverige AB Org nr: 556057-4880

H:\egna_rapp\Trafik i täta miljöer _slut.docx

TRAFIK I TÄTA MILJÖER

Svante Berglund (Modellutveckling, text och scenario) Patryk Larek (Modellutveckling, text och scenario) Joel Franklin (Scenario)

1 INTRODUCTION 6

2 INTERACTION BETWEEN MODES 7

2.1 Car on car and other motorized vehicles 8

2.2 Motorized vehicles on bike and vice versa 8

2.3 Walk on other modes 10

2.4 Shortest path and route choice by bike 11

3 TEST AREAS, CODING AND CALIBRATION 11

3.1 The model of Södermalm 11

3.1.1 Car 12

3.1.2 Bus 13

3.1.3 Bike 15

3.1.4 Cyclist behavior 17

Bicycle demand data 21

3.1.5 21

3.2 The model of Stockholm City 22

3.2.1 Model coverage 22

3.2.2 Car 23

3.2.3 Transit 23

3.2.4 Walk 23

3.3 Model calibration and modes 24

3.3.1 Behavior 24

3.3.2 Network adjustments and demand 24

4 TEST CASES 25

4.1 General overview of modeling interaction 25

4.1.1 Car 25

(2)

4.1.2 Bike 28

4.1.3 Bus 32

4.2 Car and pedestrians 33

4.3 Route 4 simulated as BRT 35

4.4 Conclusions from the tests 37

5 TRAVEL TIME IN TRANSMODELER, EMME AND GOOGLE 37

6 CONCLUSIONS 39

(3)

Sammanfattning på svenska

Täta storstadsmiljöer blir allt viktigare. En stor del av tillväxten av befolkning och arbetsplatser är koncentrerade till storstäderna medan mindre orter förlorar både befolkning och arbetsplatser. Det är rätt tydligt att en allt större del av persontrafiken kommer att ske i täta miljöer med en komplex trafiksituation. Med en komplex trafiksituation menar vi hög trängsel och en hög grad av interaktion mellan olika användningar av trafikmiljön. I dagens planeringsmodeller fångas ingen form av interaktion mellan olika färdmedel utan enbart inom biltrafiken där man tar hänsyn till personbilar och även lastbilar vid beräkningen av trängsel. Det förekommer däremot ingen försening till följd av fotgängare annat än genom klassificering av

fördröjningsfunktion.

Ansatsen projektet var att etablera en mikromodell och det har vi gjort dels inom det här projektet och dels inom ett parallellt tillämpat projekt för Stockholmsstad. Vi har dragit erfarenheter från båda som vi bakar in här.

I projektet har vi skapat en modell där avsikten är att integrera så många trafikslag som möjligt i en och samma modell och låta dessa påverka varandra inom modellens ramar.

De färdsätt vi har integrerat är: Bil, buss, cykel och gång (i City-modellen ingår inte cykel). Bussar går enligt linje i busskörfält och i blandtrafik. Cykel får röra sig på cykelbanor, cykelfält som kodats och i blandtrafik. Gångtrafikanter förekommer endast vid övergångsställen vilket ger en fördröjning för korsande trafik.

I projektet har vi upplevt ett antal praktiska problem som främst handlar om kodning av övergångar mellan länkar och fält med olika status. Sättet som cykelfält och cykelbanor ansluter till vägar varierar och är svåra att representera på ett korrekt sätt. Ofta har man gjort utformningar givet en mängd restriktioner på yta som speglar kompromisser mellan olika anspråk och ingen lösning är den andra lik. I modellvärlden är det ibland svårt att efterlikna den funktion man försökt skapa i övergångarna mellan länktyper. Det är också svårt att ställa in ett beteende hos cyklisterna som fungerar i den mångfald av utformningar som finns på Södermalm. Ofta finns små ytor att köa upp cyklister och cyklisterna tvingas ta den yta som finns oberoende av de parametrar som sätts i modellen.

En notering som kanske är självklar är att ju tätare trafikmiljö ju mer av trafiken regleras med signaler. Projektets utgångspunkt var att vi skulle studera konkurrens om utrymme men det har blivit allt tydligare att det handlar i hög grad om fördelning av tid mellan olika anspråk på en och samma yta än om att konkurrera om ytan. Signalreglering går ju ut på att temporärt reservera en yta för en trafikantgrupp. Själva trängseln på länken inom eller mellan färdmedel är en del av tidsfördröjningen men fördelning av tid i konfliktpunkterna (korsningarna) blir mer avgörande i centrala miljöer. Slutsatsen av detta är att modelleringen bör tydligare inriktas mot att studera effekten av prioritering av tid. Ett starkt intryck är att signalreglering bör hanteras explicit såsom i

mikromodeller vid modellering av centrala miljöer.

Hantering av trängsel bland bilar är väl etablerat både inom mikro som

makromodellering, cykel mindre så. Noterat är att i de områden där det är som flest trafikanter finns separata cykelbanor eller cykelfält, interaktionen mellan

trafikantgrupper är måttlig i dessa områden. Däremot förekommer interaktion där cykelbanan eller fältet tar slut (eller börjar) och dessa punkter är genuint svåra att

(4)

modellera då utformningarna inte är standardiserade. Det gäller i såväl mikro- som makromodell.

Signalreglering har en tydlig effekt på trängseln och det är att den samlar ihop cyklister i klungor som sedan släpps iväg vid grönt (detta hanteras i en mikromodell), det skapas en trängsel inom gruppen som påverkar hastigheten längs efterföljande länk. Samtidigt kan cyklister som kommer in på länken från ett annat håll som inte är signalreglerat åka ostört eftersom dom inte samlats upp på motsvarande sätt. I de här fallen är det inte alltid volymen per timme eller annan tidsenhet som är styrande utan en tillfällig ansamling efter en signal. Den dynamik som skapas skiljer också mellan snarlika länkar beroende på hur cyklisterna samlas upp (vi beskriver detta i viss detalj i

rapporten). En slutsats är att på de platser och tidpunkter där det är relevant att studera interaktion mellan fordon/trafikanter är svåra fall att hantera. Vi upplever ändå att en mikromodell hanterar det rimligt bra medan den typen av fenomen är svåra att representera i en makromodell. Även i en statisk modell där man låter volymer från cykel respektive bilutläggning påverka varandra inbördes är det ett problem. I rapporten visar vi noggrant funktionen i dessa konfliktpunkter och hur det hanteras.

Vi har genomfört en några tester eller experiment för att studera hur modellen reagerar på att man modellerar färdmedel tillsammans eller separat. Vad vi generellt kan mäta och observera i en stadsmiljö är restider där samtliga färdmedel befinner sig i trafiken medan man oftast modellerar endast bil. I praktiken har vi begränsad kännedom om restid för bil i stad som endast beror på volymen av bilar. Vi lade upp beräkningarna som så att vi kodade nätverk och signaler så gott vi kunde och använde

modellparametrar som förväntades fungera i svensk trafikmiljö. Detta gjordes för bil, cykel och buss. För fotgängare i korsningar användes räkningar där dessa var tillgängliga och annars schabloner. När modellen fungerade rimligt väl kördes ett scenario med samtliga färdmedel i trafiknätet och därefter varje färdmedel var för sig.

För bil ger modellen kring 7 % högre medelhastighet på Södermalm om man tar bort cykel och buss. I beräkningarna för cykel har vi något större variation och där ökar medelhastigheten med 5-9 %. För buss gjordes samma sak och där var effekterna något större eller hastighetsökningar kring 10-15 % när dom var ensamma i trafiknätet.

För buss gjordes också en mer detaljerad analys när vi körde linje 4 över Södermalm som BRT dvs i separata fält där det var fysiskt möjligt och under antagande om av och påstigning i samtliga dörrar. Effekten blev att genomsnittshastigheten ökade från 15 till 19 Km/h (27%) för linje 4 på Södermalm medan den sjönk från 20 till 19 för bil i linjens sträckning.

Trängsel mellan bilar upplever vi att modelleras väl i modellerna, både i City och på Södermalm. En notering vi gjort är att fotgängarvolymerna i övergångsställen har en begränsad betydelse för hastigheterna i nätverket som helhet men kan påverka volymen i enskilda svängar. Eftersom fotgängare inte modelleras explicit vid oreglerat övergångsställe eller vid sekundärkonflikt finns anledning att fundera hur detta ska hanteras i en prognossituation. I en tätare miljö kommer antalet fotgängare att öka både till följd av direkta resor med gång och som anslutning med kollektivtrafik. Ett ökat antal gångtrafikanter kommer att påverka restiden för både bil och cykel. Man kan överväga att skriva upp fotgängarvolymerna i prognossituationen eller att utnyttja de antal resenärer som väljer gång och kollektivtrafik i modellen för att beräkna troliga fotgängarvolymer. De beräkningar vi har gjort, som mer handlar om hur modellen hanterar interaktionen mellan fotgängare och andra trafikanter, ger att en ökning av

(5)

antalet fotgängare med 10 % ger en ökning av bilrestiderna med ca 0,4 %. Den restidsökningen är endast ett resultat av ökat antal fotgängare vid obevakade övergångsställen och vid sekundärkonflikter. Ökningen av förseningstiden till följd av fler fotgängare är i beräkningen ca 2 %. Slutsatsen av studien av fotgängare är att det troligen finns större problem att angripa än precisionen i antalet fotgängare i modellen.

Det här ska inte tas som intäkt för att helt avstå från att koda fotgängare i korsningar och övergångsställen vilket kan ge påtagliga effekter.

För fotgängare har vi inte tagit fram några restidsfördröjningar för än (vet inte ens om det går inom aktuell modell), vid prioritering av olika gröntid är naturligtvis dessa en post.

Under arbetet har vi jämfört restiderna i vår modell med de restider som en statisk modell genererar och med restider från Google och mätningar från Stockholms stad.

Slutsatsen av de jämförelserna är att mikromodellen verkar ha en bättre förmåga att fånga restiden i riktigt komplexa trafikmiljöer men att en statisk modell fungerar väl så bra under mer normala förhållanden. För att få mikromodellen att fungera bra krävs en hög noggrannhet avseende detaljer och finns inte utrymme för ambitiös kodning är sannolikt makro- eller mesomodeller att föredra. Vi noterade också att vår mikromodell inte är lika förlåtande avseende de trafikvolymer som modellen förses med som en makromodell och att precisionen med avseende på avresetidpunkt måste vara hög. Ett starkt intryck är att om man ska jobba regelbundet med mikromodeller (kanske även mesomodeller) bör modellernas precision ökas i båda de dimensioner som nämnts.

Vi redovisar också en ansats som gjorts i en statisk modell där man lägger stor möda på att hantera trängsel mellan bil och cykel på länk i en modell för Ottawa. Mot

bakgrund i de resultat vi noterat här tror vi inte att det är ett förstahandsalternativ för att modellera en svensk innerstadsmiljö.

(6)

1 Introduction

The planning paradigm that has guided the development of Swedish cities over the last decades has stressed the merits of high density. The observation that central areas have higher shares of both transit users and pedestrians as well as attract a broad spectrum of activities has tempted planners and politicians to extend the inner city further out beyond the most central areas. It seems that a larger share of all personal transport occurs in central areas with a more complex traffic environment. Plans and goals for the development of the transport system that are associated with these land use plans assume high shares of transit in combination with ambitious goals for share of trips by bike. Walking is already a central mode in dense areas, both as a primary travel mode as well as an access mode to transit. The high expectations on bike can be explained by a hope that bike can be a substitute for car trips.

Traffic consequences of different planning strategies have been addressed in planning literature1 and in applied regional planning2. The evaluation measure is predominantly the share of non-motorized travel, with few utility measures related to travel speed or cost. Even if plans were evaluated using e.g. a consumer surplus calculation, it is a risk that transport conditions would not be correctly captured in a travel demand model due to the nature of traffic in dense urban areas. Failure to capture traffic conditions can be expected due to the dominating use of static travel demand models that typically only capture intra-mode congestion. Intra-mode congestion (the effect of cars competing with other cars about space) is appropriate in separated traffic conditions i.e. highways but less so on inner city streets. Traffic in central areas is characterized by very different traffic conditions with a high degree of mixed traffic where cars, buses, trams, bikes and pedestrians share the same space. Such situations calls for other

approaches where intra-mode interaction can be captured.

The development towards increasing density raises some questions that will be addressed in this report.

1) To what degree it is possible to model interaction between modes and how the model will react when doing so. We hope to arrive at some recommendations if it is worth the additional work associated with modeling other modes than the mode that is focus of the study.

2) We will also study some priorities of street space.

3) We will try to shed some light on the advantages and disadvantages of dense urban structures by a more detailed calculation of travel time.

Development has been concentrated to central areas or areas bordering the city center. The goal of e.g. Stockholm is to extend the inner city by replicating the

character of central areas. There is also an ongoing densification of areas just outside the urban core with the ambition to increase the share of non-car. This calls for an approach where we model other modes with a higher ambition than we have previously.

In short, what motivates this work is the following:

1 See e.g. the literature on travel and the built environment e.g. Ewing & Cervero (2010)

2 The regional plan of Stockholm is extensively evaluated using transport models.

(7)

A larger share of all mobility will take place in urban areas with a more even mix of modes;

Interaction between modes will be more pronounced; and

Mode shares will shift away from car as a result of higher density.

Our approach is to try to capture the interaction between different modes by developing a multimodal network model where we include relevant modes in such a way that they will affect their flow across modes. We do this using microsimulation software.

Microsimulation and traffic visualization is a quite common approach to analyzing the model. A report has limitations in this respect, thus we try to use figures extensively to illustrate model behavior.

2 Interaction between modes

In this section we make a typology of different types of interaction between user groups and discuss properties of this interaction. In standard transport modeling, we usually model intra-mode interaction limited to interaction between vehicles in the road network. Standard practice in Sweden uses total vehicle volumes in static network assignment to determine link speeds. Heavy vehicles are usually not translated into car equivalents, however this is a quite simple extension that can be made³. In most parts of the road network this practice is good enough; buses with frequent stops that block cars are rare, and the influence from bikes and pedestrians is limited. For pedestrians and cyclists, the number of cars present is usually not a limiting factor for travel speed however the experience of the traffic environment may be different.

Static network assignment has proven to work fairly well in light to moderate congestion where the influence from other modes is limited. It works less well in heavily congested situations where traffic blocks upstream links, and the effect from disturbances of other modes is quite unknown.

The table below depicts all modes and directions.

Table 1. Effects we capture in our model.

Effect on:

Car Bus Walk Bike

Effect from: Car Yes Yes No, traffic signals Yes

Bus Yes Yes No, traffic signals Yes

Walk At crossings At crossings No At crossings

Bike Yes Yes No Yes

Starting with the obvious notion, we have one fundamental difference between walk and the other modes; car, bus and bike can be in the same stream of traffic while pedestrians interact by crossing the same stream. In the following sub-sections we discuss different ways to capture this interaction.

3 Given that volume delay functions have been estimated using personal car equivalents as well.

(8)

2.1 Car on car and other motorized vehicles

Congestion in the road network is the standard analysis object in traffic planning. Intra- mode interaction is well known and usually taken care of in transport modeling. In dense environments, some approaches will have limitations. The standard approach using static network assignment is suitable in most conditions. The effect of congestion on travel time in the static approach is calculated using functions relating volume to travel time (t) on the link e.g. using the BPR volume delay functions:

( ) = × 1 + ; where is passenger cars, C is capacity on the link, and are parameters.

This approach has some important limitations, one is that the dependence is within one link and if more vehicles enter the link than can exit the link, upstream spill-back will not be captured. The inability to capture spill-back is usually not a problem outside major urban areas when analyzing conditions on highways but it could be a problem in inner city areas with short links with limited storing capacity where spill-back is an issue.

The property of the intra link dependence also neglects the effect of conflicting traffic streams. Outside city areas this is a manageable problem but it will increase in an inner city area.

Low traffic volumes can be observed for two reasons: 1) low demand; and 2) high demand but insufficient capacity resulting in low volumes departing a link. The speed at a given volume can thus not be uniquely determined.

Within the static framework it is possible to take into account the effect of other modes if the effect is known. If we restrict ourselves to motorized vehicles, we know the effect on cars from other vehicles e.g. trucks and buses. This can be treated as a background flow on the link that is added to the volume on the link. The formula above will then be modified by adding the volume of other vehicles and a factor converting their

contribution to congestion in passenger car equivalents:

( ) = × 1 + ; where PCE is a factor converting other vehicles to passenger cars in terms of congestion. The effect on cars from other vehicles is simple to implement but the effect the other way around is less simple to analyze.

2.2 Motorized vehicles on bike and vice versa

There is a strong support in the literature that vehicle volumes degrade the level of service for bikes. In a literature survey of route choice models (Berglund & Engelson, 2014) the main difference found in valuation of alternative routes was not between different countries or from using different methods, but between different levels of car traffic on the route.

In a study by Broach (2012), of route choice among cyclists’ valuations of attributes in the bike network was estimated. From these estimates, relative valuations of links can be extracted. Links with high traffic is valued 1.7 times worse compared to a links with low traffic Interpreted as weights, the difference is striking; to avoid a one-kilometer link with high traffic, cyclists are prepared to take a detour of 0.7 km using links with low traffic. When considering conflicts in intersections, the relative difference is even larger for un-signalized intersections only. In un-signalized intersections, the penalty for

(9)

left/straight differs with a factor of close to 10 (see table below)! The relevance of this in a Swedish context is brought into question since intersections with volumes greater than 20 000 vehicles per day in inner city environments without traffic signals are rare.

It is worth noting that the turn penalty (expressed as an addition to non-green time) for signalized intersections is quite small.

Table 2. Interpretation of turn penalty (in km) for different types if intersections from Broach (2012).

Turn direction and traffic volume in intersection (annual average daily traffic)

Non-commute Commute

Right no signal 10 000- 0.11 0.06

Left no signal 10 000-20 000 0.27 0.15

Left no signal 20 000- 0.71 0.38

Left, straight signal 0.06 0.035

Left, straight no signal 5 000-10 000 0.11 0.06

Left, straight no signal 10 000-20 000 0.174 0.1

Left, straight no signal >20 000 1.0 0.537

From these figures, it seems important to take into account the interaction of bike with other modes. There are very few examples of this in the literature, with Gupta et.al4 as a notable exception. Gupta et.al developed a model for Ottawa-Gatineau where cross- modal impacts are taken into account. The Ottawa model uses static assignment software (EMME). Cross-modal congestion is communicated by iteratively modifying volume delay functions for both modes. Since it is an interesting approach with the ambition to capture the same phenomena we are interested in, we describe it in some detail. In the Ottawa model the link travel time function is formulated as a function of free flow, link attributes,

= ( )

where is travel time for bike, is free-flow travel time, is link delay factor, ACF is auto congestion factor =^[ ^] and ( ) is congestion effect due to bicycles. is link delay that is independent of bike flow but depends on auto flow and other attributes of the link. The auto congestion factor is the time delay on the link caused by cars moving slower than the free-flow speed. If the link is a bike path, ACF will be 1. ( ) is a standard type of vdf (BPR):

1 + ; where is = max , min + , , and

= max{ , }. is auto volume on link a and link capacity is .

The corresponding effect on car from bike uses a conversion of cyclists to passenger car equivalences:

4 http://onlinepubs.trb.org/onlinepubs/conferences/2014/ITM/Resources/26.pdf

(10)

( ) = × × 1 + ; where is passenger cars and bike volume, LF is a link factor6 and is a conversion factor from bikes to passenger cars.

In the formula, l indicates that the conversion factor could be link-specific, but, as far as we understand, a factor of 0.8 across all links is used. A cyclist is thus assumed to contribute to congestion almost as much as a car. A difference between the vdf’s is that car volumes affect the shape of the function for bikes, i.e. the contribution of another cyclist will differ depending of car volumes while bike volumes is treated as a passenger car equivalent.

This system is iterated to convergence by:

1. Assigning cars

2. Calculate bike VDF-parameters for bike using input from step 1 3. Assign bikes

4. Check convergence

a. If no, send bike volumes to car assignment and go to step 1

The sequence above is just a loop between two different assignments in consecutive steps. There are other choice models (e.g. mode) where feed-back loops loop back to car travel times.

In the equation above, ( = ( )), car volume enters three factors in different ways. It is not simple to get an intuitive feeling for the result in the end using this method. A shortcoming in this implementation does not include intersection delay, which we have previously argued is important. The framework does, however, allow turn penalties to be included to a cost of further complexity.

In applied work, this model has produced good fit to counts.

2.3 Walk on other modes

In standard transport models, pedestrians are not modeled explicitly but rather a background alternative for short trips. In Swedish models, walk and bike are modes in the demand models but these matrices are not assigned or evaluated in standard planning. From transport models, we do not have any knowledge to properly model pedestrians. There are, however, other sources of information about pedestrians such as count data used for crosswalks. Crosswalks are of two different kinds, signalized or un-signalized with pedestrian priority. Signalized intersections are pre-timed and the number of pedestrians will only cause a delay for vehicles turning right.

In inner city areas, travel for pedestrians and cyclists is usually separated physically or by a painted line, and direct interaction is typically limited to intersections. The painted line is, however, a rather weak informer and easy to miss. The degree to which this semi-mixed area affects cyclist is unclear. We have found no examples in the literature regarding the effect of pedestrians on cyclists.

6 The link factor is 1 for car links and 1.33 for links with mixed traffic.

(11)

2.4 Shortest path and route choice by bike

There is emerging literature on bicycle route choice. Important contributions have been made by Broach (2010, 2011), Halldórsdóttir et.al. (2015) and Hood (2011). A common theme portrayed in the literature is that cyclists have an aversion to sharing space with cars in dense traffic. This has been confirmed in studies by e.g. Broach et.al. (2011), Sener et.al. (2009). To capture this phenomena in a dynamic model, we must first consider what governs the behavior in a micro model. A road where bikes and cars share the same space is not a problem for a cyclist until cars are actually on the road.

The distance to surrounding cars is the problem, not a formal attribute of the link. The challenge is then to set the behavior of cyclists so that route choice is consistent with observed behavior.

This can be done by establishing parameters in the model that affect behavior in relation to other vehicles i.e. acceptance of gaps for lane changing, desired following distance and acceptable distance when stopped. These parameters can be found in the appendix.

3 Test areas, coding and calibration

This project includes two test areas. One project test area is Södermalm, which focuses on careful coding of a few corridors. The second model covers the city of Stockholm and is homogenous with regard to coverage area but lacks some detail in other respects.

3.1 The model of Södermalm

Södermalm is an area that has been previously used as a study site. The island has a mix of work places and residential areas that makes it suitable for our purposes. The geography creates distinct limits and connections with surrounding areas.

Figure 1. Area covered by the model.

0 .15 .3 .45

Kilometer s

(12)

For Södermalm, our approach was to code all modes that can possibly interact:

Car Bus Walk Bike

There are also subway lines, long-distance trains and commuter trains passing and entering Södermalm, however these trains are separated from other modes and are not modeled as vehicles. We do model passengers arriving to bus stops where some of these passengers change from trains stopping at the stations on Södermalm, but the complete chain of modes is not modeled. Passengers arriving at bus stops don not vary over the studied period and are just included to cause delay to the buses.

We should also mention that there is periodic boat traffic in the area, but we did not regard it as a relevant mode nor were the passenger flows of such a size that they could be considered of relevance.

This network represents year 2014.

3.1.1 Car

The car network is coded based on NVDB⁷ (Sweden’s national road database) as well as signal-plans delivered by the City of Stockholm. The car network is pictured below.

Figure 2. Car network of Södermalm.

Additional coding was based on aerial photographs.

7 www.nvdb.se

(13)

Figure 3. Detail of car network (Slussen before demolition).

Demand data was obtained from a traversal matrix from Sampers.

Since development of car networks for micro simulation is quite well known, we do not go into further detail.

3.1.2 Bus

For bus we initially used GTFS-data from Trafiklab8. GTFS (General Transit Feed Specification) is a data structure that provides information of all scheduled transit service. For Sweden this information is complete (at least it is for Stockholm) and contains all scheduled traffic in Sweden during the period the feed covers. Since GTFS is a standard defined by Google, it is widely utilized and Transmodeler has an import feature for it. The GTFS feed is, however, not that simple to use since every deviation from the standard route in any respect shows up as a separate route and the number of bus routes that serves the same purpose can show up as a large number of separate routes. In the end we had to do quite a bit of manual coding of the bus routes.

In order to import a GTFS feed into a Transmodeler network, the model needs to cover the whole transit route, not just what falls within the study area. Additionally, the network coding needs to be accurate enough to allow physical movement of buses.

Unfortunately, designated infrastructure for buses is not present in NVDB so this task relied on manual work. A limitation in the bus network is that there is no interaction between buses and other modes outside the study areas. The latter observation made us abandon the original network and manually code the parts within Södermalm.

8 Trafiklab.se, Trafiklab

(14)

Figure 4. Bus routes on Södermalm.

Bus routes on Södermalm are illustrated in the figure above. Buses use the same network as other modes in addition to reserved bus lanes. Traffic signals for buses were a part of the signal plans delivered by the city of Stockholm.

Figure 5. Detail with bus lane (blue) and bus stop. (Brown).

Demand for bus trips is not modeled explicitly in this model; Transmodeler only models vehicle movements not path choice of transit travelers. We are including buses

because of their contribution to congestion and to study how buses are affected by other modes.

Bus travel times, apart from the schedule, are similar to other modes determined by dynamic traffic in addition to time for boarding and alighting of passengers. Usually the number of passenger boarding will determine stop delay. Since the number of

0 .15 .3 .45

Kilometer s

(15)

passengers boarding per bus stop is potentially important data to capture bus delay time, we gathered this data9 for the most important stops.

Figure 6. Boarding passengers by stop.

In the figure above, the circles are proportional to the number of passenger boardings.

Data for passenger boardings is available for the stop area but not by direction, so this split by physical bus stop is approximate. Large numbers of arriving passengers can be found where subway lines intersect with the major bus lines. We also have two

commuter train stations where several bus lines pass by.

3.1.3 Bike

Coding and setting the behavior of cyclists was the most challenging element of this project and the reasons for that can be attributed to:

1. Network issues 2. Behavior of cyclists

In a mature city like Stockholm, the space available for traffic is limited by the distance between blocks. This space has since 196710 been used for cars, parking and

pavements. Since planning for bicyclists has only recently become an issue, this mode has been given the leftover space, and solutions for bikes have been governed by what is possible within the leftover space.

Comprehensive bike networks consist of bike paths, bike lanes and lanes with mixed traffic. A few examples are provided below.

9 Source: SL

10 In 1967 Sweden switched to right hand traffic and trams were abounded in Stockholm.

(16)

Figure 7. Example of network, bike paths/lanes in green.

A network with bike paths is illustrated above shown by the lanes in green.

In the figure below (from Hornsgatan) are a few examples. Following this street, we can note that we have separate bike paths, bike lanes and bike in mixed traffic. On some segments the bike lane is to the right, while on others it is between car lanes and parking to the right (not shown in figure).

Figure 8. Example of the position of bike lanes.

Route continuity will of course be suffering and how this is represented in the model is important to the model behavior. The challenge is where the route changes character, going from separated to mixed or vice versa. Mixed traffic on continuous links is not that complicated compared to these areas. These areas do not follow a standard manual, as we mentioned, and several of these “intersections” are unique.

(17)

Figure 9. Detail of bike network at Hornstull

The figure above illustrates one of Stockholm’s more complicated traffic solutions that includes bikes. The number of cyclists travelling from the south to north direction is approximately 1200 during the morning peak hour (one-hour during October). At these intersections, we experimented with the coding to avoid a model behavior where cyclists did not completely block conflicting streams of cyclists. We also needed

enough storage space for stopped cyclists at the traffic signals. In some cases we were forced to add an additional bike lane and in other cases we widened the lane available.

Cyclists in the model use the space coded as lane and intersection area, while in reality cyclists use the needed area regardless of status (e.g. intended for pedestrians. In this case we are not clear if the source of the problem is behavior or network coding or a combination where the environment forces a local behavior.

3.1.4 Cyclist behavior

Cyclists in the model have defining properties similar to cars, such as ratios of power to mass, sizes and weights, and so on. These parameters have not matured in the simulation literature and properties of cyclists differ considerably between countries.

Examples of this include queuing behavior11 and respect for rules, for example how a queue will be built up at a traffic signal and how this queue will be discontinued during the green phase. In contrast to cars, cyclists can pass another cyclist in the same lane if it is wide enough and the distance is acceptable.

Without manipulating the model and network, an ordinary queuing behavior similar to a line of cars will be formed. An orderly queue will be lengthy and prevent all cyclists from passing through during one green phase. There exists a strong incentive for cyclists to form a wide and compact bunch that allows all or most cyclists to get through during the next green phase. To obtain a realistic behavior of this phenomenon, we decreased the acceptable distance between cyclists at stops and artificially widened areas close to

11 There has been intensive debate about cyclist behavior in Stockholm in the press.

The discussion focus on segments of the population that have an aggressive behavior in the traffic. It is possible that future modeling attempts will benefit from segmentation in to driver groups.

(18)

signals. The change of the acceptable distance between cyclists did however, cause blocking of cyclists with crossing paths. This may need some additional work to find a better balance between different behaviors.

Figure 10. Example of coding and “queuing” behavior of cyclists. Picture of Götgatan towards Slussen.

After some trial and error, we obtained a behavior in the model that corresponded to observed behavior.

Observed behavior was collected from previous fieldwork (films, photos, counts and observations).

Congestion between cyclists shares some but not all of the same characteristics as with cars, for example the behavior mentioned above in association with stop signals.

The formation of bunches at stop signals will cause local congestion along downstream links even if the average hourly number of cyclists does not motivate congestion. This is the case along Götgatan, Where overtaking cyclists will experience congestion to some degree that originates from stops. According to counts made by the police, bunching at Götgatan towards Slussen (the previous picture) could consist of 25-35 cyclists, and, in addition to congestion, the formation of bunches highly increases the risk of accidents12.

12 The reason for the police to count cyclists at these spots.

(19)

Figure 11. Example of “local congestion” created by a stop signal.

The figure above illustrates the effect of the stop light shown in the lower part of the figure and the corresponding congestion shown in the upper part of the figure. These bunches will stretch along downstream links if there is not another stop signal, often the case in urban traffic.

As a consequence of traffic signals packing cyclists together, the level of congestion along a link is dependent from the direction a cyclist enters the link. In the figure below we have cyclists on the northbound link (marked by the red circle) travelling from the east incoming link with limited demand. This is in contrast to what will be experienced by the cyclists waiting on the south incoming link.

(20)

Figure 12. Example of direction dependent congestion.

We should also note that vehicles (buses) and bikes travelling in the same direction are usually in the same signal phase and thus both modes are grouped together and will possibly contribute to the experienced congestion. The consequence of a common signal phase is that it will be difficult to use the outside lane to pass slower cyclists.

These mechanisms may deserve further interest and could be of interest for dimensioning.

These kinds of phenomena cannot be captured by static models using volume delay functions since delay is based the average volume during a period on one link

regardless of temporal variations. Delay due to temporal accumulations of vehicles can be captured, at least to some degree, by micro models. A static model will thus over estimate travel time for cyclists coming in from the right direction (in the figure above) and under estimate the travel time for cyclists entering the intersection from the south.

Due to the nonlinear nature of volume to time functions this errors can be quite large.

(21)

3.1.5 Bicycle demand data

Demand modeling for bicycles has not gained much interest until recently; however, there is a strong tendency for change in that respect¹³. We started by assigning a standard demand matrix developed in a previous project not specialized for bike (Berglund & Engelson 2014) to a bike network and compared the results to bicycle traffic counts. In the referred work, we did not calibrate the matrix to any traffic counts but the accuracy of the numbers surprised us. The results were encouraging, so we decided to continue using a matrix from that model and extracted a traversal matrix for our study area.

The input matrix for a simulation project is typically a modeled matrix that has been modified with respect to traffic counts. Adjustments of bike matrices poses a few problems that are less pronounced in analyses of car traffic. Bicycle data has very high variability due to weather, season and time of day. During winter months, the mode share for bike is limited while the opposite is true during the warm period of the year.

Furthermore, there is quite strong variability between years in the number of cyclists, for example if it was a late versus early spring, and so on.

Over the course of a day, bike has more pronounced peaks compared to car (see figure below). The figure below is based on data from one of the fixed data collection points that is in operation all year. The reason for the concentrated pattern of demand can be explained by homogenous trip purposes of cyclists (work and school). If we look at the time distribution of work trip with cars it is quite similar to the one with bike. The time distribution of cars obtained from traffic counts depends on other purposes spreads the peak. The effect of concentrated demand is that cyclists will contribute to congestion during a short period of the day, but that period is during peak for all modes.

The result will be that most cyclists will experience a problematic traffic situation and that cyclists will contribute to congestion during the period when space is most scarce.

Figure 13. Time profile for cyclists passing Skanstull during a Tuesday in October 2013.

13 There are ongoing projects in Sweden to develop models for bike and a variety of approaches under testing abroad.

0 50 100 150 200 250 300 350 400

00:00 04:48 09:36 14:24 19:12 00:00

Cy cl is ts p er 1 5 m in

Time

Skanstull

Norrgående

Södergående

(22)

Strong differences in demand over time, as we have for bike, are a challenge for a model. Volumes at different measurement points or times are hard to get correct and limited errors in the model or underlying data reflect poorly on the model.

3.2 The model of Stockholm City

To illustrate some properties of the model, we used a model of Stockholm inner city as another area. We used this for two reasons: the model area was available and using two areas supports add some insight.

3.2.1 Model coverage

The figure below depicts the model network for Stockholm’s inner city area.

Figure 14. Picture of the network in the city model.

For the model of Stockholm City we coded:

Car Bus Tram Walk

As noted above, bike is not present in the City model. Bus is coded using headways, not according to the actual schedule that is a simplification that is regarded to not affect the results. In the City-model we have also included a short tram line.

(23)

3.2.2 Car

The car network is coded based on NVDB as well as signal-plans delivered by the City of Stockholm. We use the same level of detail in the city model as we have for the model of Södermalm.

3.2.3 Transit

The transit network in this city model consists of buses and a short tram line. We only include transit vehicles in this application to affect the travel time by car.

Figure 15. Picture of transit network in the city model

3.2.4 Walk

Pedestrians have effects on other modes for example on secondary conflicts at right turns, at un-signalized crosswalks, and indirectly at signalized crosswalks. In the city center we have access to quality pedestrian counts at the intersections indicated in the map below.

(24)

Figure 16. Intersections with pedestrian counts.

3.3 Model calibration and modes

Both models were calibrated separately and with different data. The city model that was developed for Stockholm was a bonus to this project. The network data and signal plans for both models share a common source and thus should be of similar quality.

3.3.1 Behavior

Behavioral parameters are the same in both models with the exception of bike

parameters that are not present in the city model. Parameters of driving behavior were developed by ÅF (2016), see also appendix 2. Parameters for bike were adjusted to provide a better correspondence with observed behavior (see appendix 2). Parameters that were adjusted for bike include the distance between stopped bikes and the

maximum weight of cyclists (sum of weight of cyclist and bike).

3.3.2 Network adjustments and demand Our calibration process followed these steps:

1. Assign an un-calibrated matrix to the network for each mode separately.

2. Simulate the traffic and correct problems related to network coding errors.

3. Make a second assignment continuing with each mode separately. The reason to initially treat modes separately is purely practical- to identify the worst problems and get a complete model run.

4. Check link volumes and travel times and adjust the demand matrix.

5. When each mode assigned separately is acceptable with regard to volume (and speed14), proceed to assign all modes simultaneously to the network.

6. Check, adjust and correct demand and further network problems.

14 This is of course unknown since what we observe is the traffic we have that is mixed.

(25)

This short numbered list represents a lot of work, some of which is mentioned in other sections.

4 Test cases

In this section we make five different tests and analyze the outcome of these tests. We initially try to provide an overview of the model behavior and the effect of modeling all modes, both together and separate.

It is common to model car traffic as a separate mode, which is of course no problem in most situations where the traffic is actually separated. Furthermore, simulations of urban traffic are usually done for car-only assuming that the contribution of non modelled modes on congestion is limited or a fixed disturbance on the link. Thus, the calibrating is done for a situation with all modes where the resulting delay in travel time is attributed to intra-mode congestion and signals.

4.1 General overview of modeling interaction

We use a reasonably well calibrated model and begin with the current situation that constitutes our reference alternative. Our reference scenario is a model with all modes (that we model) present i.e. car, bus, bike and pedestrians at crosswalks. In a series of model runs, competing traffic is removed, modes are simulated one at a time, and then outcomes are checked. Note that the alternative scenario is what usually is done, i.e.

assigning each mode separately, and our reference scenario is an extension of standard modeling praxis.

These experiments are hard to validate since we cannot observe an experimental situation where modes are isolated in urban environments. On the other hand, validation of single mode urban traffic is difficult to validate since single mode undisturbed urban traffic is also hard to find. All traffic we can observe is a result of interaction regardless of whether we model this interaction or not. What we learn will be the influence of each mode on travel time.

One restriction is that if we do not change the traffic signals, our scenario will only measure inter-mode congestion, not what can be achieved by a total separation of modes. This is partly intentional and partly done to limit the amount of work.

Since simulations are random processes, we can draw limited conclusions from a single simulation. We make 10 simulation runs in batch form and take averages across these runs. We simulate the period from 8.00 to 9.00 and include a cool down period until 9.30. Results are presented for trips ending in ten-minute intervals. Congestion builds up during the first 10- 20 minutes, so representative conditions will be obtained during the last period of the studied hour. We will study average speed and hours of delay in the model. Delay is the difference between free-flow travel time and travel time in a loaded network.

4.1.1 Car

First, we study car traffic in the reference scenario with all modes along with car in a simulation where we have excluded all other modes.

(26)

Figure 17. Comparison of speed in a model run with all modes and with cars only.

In Figure 17 we compare speed in the two situations. Average speed stabilizes around 25-30 km/h with a difference of 1-2 km/h. The difference remains in this range across the studied period.

Figure 18. Increase of travel speed with car in a separated network.

In Figure 18 we present the relative difference between the two situations. In this model run, the speed for other modes seems to increase with 3.5 % to 7 %.

In Figure 19 we report delay for one category of cars (where there is no significant difference between different types of cars) and the relative delay is shown in Figure 20.

0.0 5.0 10.0 15.0 20.0 25.0 30.0 35.0 40.0 45.0

08:10 08:20 08:30 08:40 08:50 09:00

Average speed

Intervall ending Car all modes Car separated

0%

1%

2%

3%

4%

5%

6%

7%

8%

08:10 08:20 08:30 08:40 08:50 09:00 Interval ending

(27)

Figure 19. Compared delay in hours in a model run with all modes and modes separated for car drivers.

Figure 20. Percent change in delay for car drivers.

With the exception of one interval, the delay seems to be just abowe 8%.

The effect on car from bike, bus and pedestrians in these model run is a decrease in speed between 3.5 - 7 % with a total delay of approximately 8 %.

Comparison with the static approach

Since it will be difficult to measure the impact of bikes on cars in a real world experiment we can compare with other attempts to capture the same effect. In the previously referred work from Ottawa, volume delay functions were used where the effect of bikes was included as PCE (passenger car equivalences). One bike equals 0.8 cars and the argument for the relatively high value relative to their physical dimensions is that bikes move slowly and thus got a quite large impact on the traffic flow.

0 5 10 15 20 25 30 35 40

08:10 08:20 08:30 08:40 08:50 09:00

Delay hours

Intervall ending Car all modes Car separated

-12%

-10%

-8%

-6%

-4%

-2%

0%

2%

08:10 08:20 08:30 08:40 08:50 09:00

Intervall ending

(28)

Figure 21. Travel time for car under different volumes for bike in mixed traffic.

In our application, there are bike paths or lanes along all major bike routes and the impact of large bike flows will be in crossings and on limited segments. There are however links with mixed traffic were there could be an impact. In Figure 21 we illustrate a vdf used in Ottawa with different number of cyclists. On an inner city street with roadside parking, 300 cyclists per hour will increase car travel time about 13 %.

Due to the nonlinear form of the vdf’s the contribution from bikes will be dependent on car volumes as well. In the example, we used 700 cars per hour. On a more heavily congested street, the effect will be stronger, but most likely below an increase of car travel time of 25 %. As we mentioned only a limited number of links with mixed traffic will have large bike volumes and the effect from bikes on cars would with regard to the functions used in Ottawa be in the lower range.

In our application, we also included buses and we treat these volumes in the same way as bikes, i.e. using PCE. Buses are present on a limited number of links, some with heavy traffic.

4.1.2 Bike

In this scenario, we remove all cars and buses and check the model behavior on cyclists. As previously mentioned, modeling cyclists is a bit of a challenge.

0 2 4 6 8 10 12 14

0 500 1000 1500 2000

Min

Car volume

Travel time car

Bike(100) Bike(300) Bike(600) Bike(1200)

(29)

Figure 22. Travel speed compared for bike in a model run with all modes and one with just bike.

We note that the speed, around 19 km/h in mixed traffic, in our simulations may be a bit overstated. This may require some adjustment in future work. There are several

measurements available ranging from pure amateur measurements to more organized attempts. In Copenhagen15 an average speed of 15.5 km/h was reported, however Denmark is a more mature bike country where a larger share of the population cycle in comparison to Sweden where cycling is more of an enthusiast mode. A measurement by NTF16 (an organization for traffic safety) measured cyclists in Lund using the same equipment as the police use. On bike paths, the speed was 23 km/h and in the car lane 36 km/h. There exists high variability in all bike speed measurements we have seen, with bike speed usually in the range of 10-35 km/h.

In our simulations we have a free-flow travel speed of close to 24 km/h, mixed traffic speed of 19 km/h, and 20 km/h when we simulated bikes separately. Mixed traffic thus contributes to a difference of between 2 and 8% in the simulations. The relatively limited effect of simulating mixed traffic can be discussed. One possible explanation is that traffic signals picks up most of the interaction between modes. In our scenario without cars, we have kept the signals unchanged. It is also possible that we have underestimated the interaction by coding away model artifacts at problematic

intersections; it is possible that we have made it too easy in the model. We have also mentioned the fact that there are bike lanes in most important streets and that will reduce congestion to be within one mode.

15"Bicycle statistics". City of Copenhagen website. City of Copenhagen. 13 June 2013.

Retrieved 12 December 2013.

16 http://www.ntf.se/sydost/default48270.asp 0

5 10 15 20 25

08:10 08:20 08:30 08:40 08:50 09:00

Average speed

Interval ending Bike all modes Bike only

(30)

Figure 23. Increase of travel speed with bike in a network with no cars.

Comparing estimates of delay with and without mixed traffic, the total delay decreases by about 10% in a scenario without competing traffic.

Figure 24. Compared delay in hours for cyclists in a model run with all modes and modes separated.

0%

1%

2%

3%

4%

5%

6%

7%

8%

9%

10%

08:10 08:20 08:30 08:40 08:50 09:00 Time

0 10 20 30 40 50 60 70

08:10 08:20 08:30 08:40 08:50 09:00

Delay hours

Time Bike all modes Bike only

TRAFIK I TÄTA MILJÖER Svante Berglund (Modellutveckling, text och scenario) Patryk Larek (Modellutveckling, text och scenario) Joel Franklin (Scenario)

TRAFIK I TÄTA MILJÖER

CONTENTS

1 INTRODUCTION 6

2 INTERACTION BETWEEN MODES 7

3 TEST AREAS, CODING AND CALIBRATION 11

4 TEST CASES 25

5 TRAVEL TIME IN TRANSMODELER, EMME AND GOOGLE 37

6 CONCLUSIONS 39

1 Introduction

2 Interaction between modes

2.1 Car on car and other motorized vehicles

2.2 Motorized vehicles on bike and vice versa

2.3 Walk on other modes

2.4 Shortest path and route choice by bike

3 Test areas, coding and calibration

3.1 The model of Södermalm

00:00 04:48 09:36 14:24 19:12 00:00

Cy cl is ts p er 1 5 m in

Time

Skanstull

Norrgående

Södergående

3.2 The model of Stockholm City

3.3 Model calibration and modes

4 Test cases

4.1 General overview of modeling interaction

Travel time car