A JOINT ROUTE CHOICE MODEL FOR ELECTRIC AND CONVENTIONAL CAR USERS

Transcription

1 A JOIN ROUE CHOICE MODEL FOR ELECRIC AND CONVENIONAL CAR USERS Anders Fjendbo Jensen, homas Kjær Rasmussen DU Management Engineering, echnical University of Denmark, Kgs. Lyngby, Denmark {afjje, Carlo Giacomo Prato School of Civil Engineering, he University of Queensland, Brisbane, Australia, Abstract Understanding the behaviour of electric car (EV) users is important in a number of ways. Beside potential environmental effects, there is a need to understand other related effects, such as effects on the electricity network and the transport network. he objective of this study is to use revealed preferences (RP) data to investigate differences in route choice behaviour between conventional car (CV) and EV users. A joint mixed path size correction logit model, including free flow travel time, trip length, number of left turns and number of right turns is presented. For all attributes the EV parameter has a higher negative value than the CV parameter, of which the highest difference is for the trip distance. Furthermore, several interactions with household characteristics and temporal characteristics such as weather and time of day are tested. One conclusion is that EV users are more sensitive to trip disturbances and trip length than ICV users, even after taking into account scaling differences.

2 1. Introduction Worldwide, governments have committed to reducing air pollution and carbon emissions. With a higher share of renewable sources in the electricity production, battery electric cars (EVs) could play a significant role in maintaining these commitments. However, the impact of a large-scale introduction of EVs is not obvious and there is a need for a better understanding of the behaviour of EV users in order to determine effects on environment, the electricity network and the transport network. Potential EV users benefit from an increasing availability of EV models, with greater comfort and better driving performance. Furthermore, due to a simpler and more efficient drivetrain, EVs have the potential to be cheaper to run and maintain compared to CVs. o obtain these benefits, however, the EV users must presently accept that the car has a limited driving range between charges and that charging of the battery, depending of the available facilities where the car is parked, takes between 20 minutes for recharging up to 80% of the capacity and usually several hours to reach full capacity. hus, there are limits to the travel that can be performed with and EV, and for most car users, it would not be possible to exchange their current CV with an EV without some level of adaption. Both these benefits and the limitations might therefore lead to behavioural changes (Jensen & Mabit 2015). EV users might avoid longer and less-planned trips and, when deciding on a route, they might select roads where the general speed is lower, the trip length is shorter, or the charging facilities are better. On the other hand, over a longer period of time, many users do not need charging other than overnight charging at home in order to keep up with their current behaviour (Christensen et al. 2010). Growing literature shows an increasing interest in EVs and their market, but only few studies have been based on actual EV usage, due to lack of data. Current EV travel demand studies are usually based on data collected from users of conventional gasoline or diesel engine cars (CVs) (see e.g. (Golob and Gould 1998; Pearre et al. 2011; Greaves et al. 2014). he objective of this study is to use revealed preferences (RP) data to investigate differences in route choice behaviour between CV and EV users. o our knowledge, this is the first time that a state-of-the-art route choice model has been estimated on RP EV data. In addition, the level of detail in the data allows for accounting for topology, weather and socioeconomic background.

3 2. Method he analysis conducted in this paper is based on several sources of data and a long range of data processing was needed. rip information was obtained from a large-scale EV demonstration project in Denmark from 2011 to he dataset consists of detailed GPS traces of each trip done by EVs and CVs as well as information about each household conducting the trips. In the processing of the data, each trip was furthermore associated with weather data from the nearest weather station at the time and location of departure. o facilitate the estimation and allow the inclusion of networkspecific attributes, the GPS traces were then map matched to a detailed network including all major and minor roads in Denmark, and alternative choices were generated by an advanced choice set generation process. he final model is estimated jointly on EV and CV data so that it is possible to evaluate consistency in the effect of each variable across vehicle technology. 2.1 Sample he GPS traces from travel in EVs and CVs was collected from 100 households in 13 different municipalities. An EV was available in a household for three months and then the car was redistributed to another household. his means that the trial study covered all seasons, but each household would not. he households all owned a CV prior to the study and data was collected from this vehicle one month before and one month after the household received the EV. Each participating household borrowed either a Peugeot ion, a Citroën C-Zero or a Mitsubishi ImiEV, which are all similar small sized cars with room for 4-5 persons and limited luggage. he NEDC standard driving range is for all cars 150km (Peugeot 2017), but according to (Fetene et al. 2016), only 7% of the trips could achieve this based on the energy consumption measured. On average, the driving distance was less than 90 km. All households got a home charger installed, but beside this they had access to public charging infrastructure covering most parts of the country including 3.7kW AC chargers and 50kW DC chargers. he households did not have to pay for the EV or the home charger, but they had to pay for the electricity consumed during the trial. When the household received the EV, they were encouraged to use it as their primary car. Participation in the trial was based on a voluntary online application process from which household information was collected. In order to participate, the households needed to already own at least one car and have access to a dedicated parking space where the home charger could be installed. From those who successfully fulfilled the criteria, the project managers selected the test households, based the provided background information with the clear intention of representing a broad range of the Danish population. Please find a description of the sample in able 1. able 1: Sample description Sample V# Variable Description N Mean S1 male Household head is male S2 ageloq1 Household head is below 38 years old S3 ageupq3 Household head is above 52 years old S4 city Household is in a city area S5 highedu Household head has a bachelor s degree or above S6 highinc Household head's income is above DKK

4 2.2 Map Matching he GPS traces were matched to the very detailed NAVEQ street network (NAVEQ 2010), which consists of 636,243 links covering the entire country and all road classes from large highways to minor local roads. he high level of detail of the network is crucial, as EV users might use smaller roads with lower speeds in order to save energy due to current technological restrictions on driving distances. After the map matching a filtering process was applied. his removed observations with a length of less than 2 km due to the likely problem of generating relevant alternatives, as well as observations in which more than 20% of the route were generated as shortest-path searches between segments to which GPS data could be matched (gaps in GPS traces). he map matching and subsequent filtering induced a dataset consisting of observed routes for 8968 EV trips and 6678 CV trips from 107 households. he routes from the CV were observed in the month before and after the EV was received while the routes from the EV were observed for the entire three month period of the trial. Each observation has been supplemented with weather information from local weather stations, inducing that information about precipitation, wind speed, temperature and visibility at the time of departure is available. Figure 1 illustrates the spatial distribution of the observed routes. As can be seen, the 13 municipalities are distributed across the whole country, covering a large part of the network. As seen in table 2, trips done by EVs are shorter in time and length than trips done by CVs Figure 2: Spatial distribution of observations. Amount of observations (EV or CV) on links

5 able 2: rip characteristics CV EV V# Variable Description N Mean N Mean A1 ffttchosen Avg. Free-flow time of observed trip (min) A2 lengthchosen Avg. Length of observed trip (km) A3 ltchosen Avg. Number of left turns of observed trip A4 rtchosen Avg. Number of right turns of observed trip Z1 is_weekend_1 Observed trip departs on a weekend Z2 peak_morning Observed trip departs during morning peak Z3 peak_afternoon Observed trip departs during afternoon peak Z4 period3 Observed trip departs one month after EV is received Z5 rainupavg Precipitation at day of departure is above 0.09 mm Z6 sunupq3 Sunshine at day of departure is above 36 min Z7 temploq1 Avg. temperature at day of departure is below 5.1C Z8 templomedian Avg. temperature at day of departure is below 9.7C Z9 windupq3 Avg. Wind speed at day of departure is above 5.1m/s Choice Set Generation Following the procedure in Prato et al. (2014), route choice behaviour is modelled with a two-stage approach consisting of choice set generation to generate alternative routes to the observed, and then a subsequent model estimation. he first stage used a doubly stochastic generation process to generate a choice set consisting of a maximum of 100 unique alternatives for each observed route. Length and free-flow travel time were considered, and the doubly stochastic approach allows considering that drivers might perceive costs with error and that perceptions might differ between drivers. he distributions of the parameters and the error term were adopted from the Danish National Model. Subsequent to the generation of the alternatives, the dataset were filtered to exclude observations for which the choice set contained only one alternative route or did not contain any alternative reasonably similar to the observed route (maximum overlap in length less than 80%). he latter resulted in the removal of approximately 25% of the observations for both EVs and CVs (Figure 2). Overall, the filtering resulted in 8962 and 6678 observations to be used in the model estimation for EVs and CVs, respectively.

6 Overlap hreshold 1 0,9 0,8 0,7 0,6 0,5 0,4 0,3 0,2 0,1 0 0,00 0,10 0,20 0,30 0,40 0,50 0,60 0,70 0,80 0,90 1,00 Cumulative Share of observations CVs EVs Figure 3: Coverage distribution As a final step before the estimation of the route choice model, the following attributes were calculated for each alternative in the choice sets for all observations; (i) free-flow travel time, (ii) cost, (iii) number of left turns, (iv) number of right turns. Ideally, the attributes should also contain congested travel time, as this is known to highly influence route choice. However, given the extent and disaggregate level-of-detail of the network, no information of congested travel times in across different time periods of the day were available; (i) congested travel time cannot be retrieved from any existing traffic assignment model as none exists which operate on such a detailed network, (ii) no GPS dataset were available which is sufficiently large to give reliable link travel times across several periods of the day on the network used. 2.4 Route Choice Model In the second stage, a mixed path size correction logit model (MPSCL) was estimated for modelling route choice behaviour (Bovy et al. 2008; Prato et al. 2014). Comparison of EV and CV preferences is made possible by estimating jointly across data from each technology using a logit scaling approach with at least one generic parameter across data (Bradley & Daly 1997). For each observation n, a linear-in-parameter utility function Vnj, was specified for each route j within the choice set to estimate a PSCL model (Bovy et al. 2008): V EV kj = β EV X X EV j + β EV XS X EV k S k + β EV XZ X EV j Z j V CV kj = θ CV (β CV X X CV j + β CV XS X CV k S k + β XZ CV X j CV Z j ), (1) where: Xj = attributes of the route, Snj = Socio economic characteristics of the household k, Zj = temporal characteristics of the route e.g. weather at time of departure, θcv = scale applied to CV data and βx, βxs,βxz = parameters to be estimated.

7 he attributes X available for the routes were free-flow time, length, number of left turns and number of right turns. A path size correction psc j captures the similarity across alternative routes within choice set C n and is defined as (Bovy et al. 2008): psc j = L a L ln δ aj, j a Γ j Where L j is the length of route j, L a is the length of link a, Γ j is the set of links belonging to route j and δ aj is a link-path coincidence dummy which is equal to 1 if link a belongs to route j and 0 otherwise. he probability of selecting the observed route i in the choice set C n is equal to: j C n P i = ex p(v ki j C n ex p(v kj + β psc psc i ) + β psc psc f(β θ)dβ, j ) (2) where denotes the vehicle technology (EV or CV), psc j is the path size correction and β psc is a parameter to be estimated. he βs from equation 1 are random parameters distributed with probability density function f(β θ)dβ in order to account for preference heterogeneity across households. he probability for each household k to select route j is then integrated over the distribution of the betas. Due to this, the probability does not have a closed-form expression and thus the maximisation of the log-likelihood function is based on simulation: N J SLL = d ni l n { 1 R [ n=1 j=1 R r=1 j C n V ki,r ex p(v kj,r + β psc psc i + β psc psc ]}, j ) Where SLL is the simulated log-likelihood, N is the number of observations, J is the number of alternative routes, dni is a dummy that is 1 if route i for observations n was the selected route and zero otherwise and r is one of R draws used in the simulation. We assume that the choices are correlated across households. hus the parameters are restricted not to vary across different observations of the same household. hey were estimated by using 100 random draws from a modified Latin hypercube samling method in BIOGEME (Bierlaire & Fetiarison 2009). 3 Results able 1 presents the parameters of the joint path size correction logit model. Initially, a base model with the four main attributes were estimated, all with random parameters with a normal distribution. echnology specific parameters were estimated for all attributes except for free-flow time as at least one parameter must be held generic across the two sets of data. he base model therefore consists of 7 random parameters. hen for each of these four main attibutes, interactions with each of the 6 household characteristics and 9 temporal characteristics of the trips were testes, i.e. a total of further 60 models were estimated. Finally, all significant interactions were added to the base model After a few rounds of adjustments, the final model with 28 parameters was obtained. he results of the final estimation is presented in able 4 below.

8 able 4: Parameters of the mixed path size correction logit model estimated jointly on CV and EV data. Value Free-flow time Standard deviation of free-flow time Free-flow time * city Free-flow time * highedu Free-flow time * male Free-flow time * rainupavg Generic CV EV Robust t- Robust t- Value Value test test Robust t- test riplength Standard deviation of triplength riplength * peak_morning riplength * sunupq Number of left turns Standard deviation of number of left turns Number of left turns * city Number of left turns * templomedian Number of right turns Number of right turns * city Path size correction Standard deviation of number of right turns Scale parameter for CV trips Number of MLHS draws 100 Log likelihood Number of estimated parameters 28 For all main attributes, a significant parameter with the expected signs was obtained. Furthermore, It is seen that for trips in households located in a city, trips, where the household applicant was male and for trips conducted on a day with higher than average rain, free-flow time seems not to be as important (positive parameter) as for the remaining trips. On, the other hand, trips conducted in high-education households, free-flow time has a negative parameter (probably due to correlation with income). Interestingly, we see that the parameter for trip-length is much larger for EV than for CV. his clearly shows that there is a difference in behaviour for the two technology types. Moreover, if the trip is conducted during the morning peak, the negative effect of length is reduced for EV, which could because this is the time most individuals goes to work and these types of trips are easier to schedule. It seems that EV users seem to avoid left and right turns more than CV users if the household is located outside a city. If the household is located in a city, the effect is similar across technologies (the sum of the main parameter and the interaction with city).

9 References Bierlaire, M. & Fetiarison, M., Estimation of discrete choice models: extending biogeme. In Swiss ransport Research Conference (SRC),. Ascona. Bovy, P., Bekhor, S. & Prato, C., he Factor of Revisited Path Size: Alternative Derivation. ransportation Research Record: Journal of the ransportation Research Board, 2076, pp Available at: Bradley, M.A. & Daly, A.J., Estimation of logit choice models using mixed stated preference and revealed preference information P. Stopher & M. Lee-Gosselin, eds., Oxford: Pergamon. Christensen, L., Kveiborg, O. & Mabit, S.L., he Market for electric vehicles - what do potential users want. In 12th World Conference on ransportation Research. Fetene, G.M. et al., Harnessing big-data for estimating the energy consumption and driving range of electric vehicles. In he 95th ransportation Research Board (RB) Annual Meeting, January 10-14, Washington D.C., USA. Golob,.F. & Gould, J., Projecting use of electric vehicles from household vehicle trials. ransportation Research Part B: Methodological, 32(7), pp Greaves, S., Backman, H. & Ellison, A.B., An empirical assessment of the feasibility of battery electric vehicles for day-to-day driving. ransportation Research Part A: Policy and Practice, 66(0), pp Jensen, A.F. & Mabit, S.L., Modelling real choices between conventional and electric cars for home-based journeys. In he 14th international Conference on ravel Behaviour Research (IABR), July 19-23, Windsor, United Kingdom. NAVEQ, NAVSREES Street Data Reference Manual v2.8, Pearre, N.S. et al., Electric vehicles: How much range is required for a day s driving? ransportation Research Part C: Emerging echnologies, 19(6), pp Peugeot, Peugeot homepage. Available at: door/technical-information/ [Accessed January 25, 2017]. Prato, C.G., Rasmussen,.K. & Nielsen, O.A., Estimating Value of Congestion and of Reliability from Observation of Route Choice Behavior of Car Drivers. ransportation Research Record, 2412(2412), pp