You are on page 1of 6

2016 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC)

June 1 -4, 2016, Busan, Korea

Hazard Based Modelling Of Electric Vehicles


Charging Patterns

Nicolo Daina and John W. Polak


Centre for Transport Studies
Imperial College London
London, United Kingdom
n.daina@imperial.ac.uk

A bstract-Understanding and predicting charging behaviour implemented in predictive models is essential, but despite
of electric vehicles' users is essential for the appropriate design of considerable research efforts have been carried out, work
charging services and for the implementation aggregator services towards this goal is still needed. So far charging behaviour
mediating between electric vehicle drivers, electricity markets, research has been mainly either simply descriptive (e.g. [8]),
distribution system operator, and transmission system operator. aimed at testing psychological models unsuitable for
Research into actionable charging behaviour insights to be implementation for quantitative prediction [9, 1 0], or when
implemented in predictive models has so far been modest. The
aimed at charging choice behaviour predictive models reliant
present paper intends to contribute at the development of
on choice experiments based on hypothetical choice situations
predictive models of charging patterns for operational
around future smart charging scenarios (e.g. [1 1 , 1 2]). One
implementation. It proposes hazard-based analysis of the gap
notable exception to this trend is the work by Zoepf et al who
times between charging events model with time dependent
use real word charging and driving data from monitored
covariates in order to investigate which set of potential covariates
that will enable an eventual formulation of short-term predictive
vehicles to estimate mixed logit for charging occurrence at the
model of the timing of charging events. Empirical estimation of
end of a trip, using as explanatory variables the state of charge
the hazard model shows that both monitored vehicle state at the end of a trip and travel patterns 'attributes (e.g., trip
variables (e.g. state of charge, cumulative average driving speed) distance, time to next departure, trip end time) [1 3].
and individual characteristics significantly affect the Additionally, the inclusion of insights from psychological
instantaneous rate of occurrence of charging events. studies of charging behaviour into models for quantitative
prediction is attempted by Daina et ai, who model drivers'
Keywords- electric vehicles, electric vehicles charging typical charging frequency and state of charge (SOC) before
behaviour, charging behaviour prediction, hazards models. charging as function of both their typical travel patterns and
perceived charging motives [1 4].
I. INTRODUCTION

It has long been accepted that deployment of electric


A. Aim and Objectives
vehicles (EV) powered by electricity from low carbon The aim of this paper is to further advance the actionable
emissions grids can provide significant benefits in terms of understanding of EV users charging behaviour for the
reduction of the climate impact from transportation and development of short-term predictive models to facilitate the
reduction of its reliance on oil-based fuels [ 1 ]. participation of charging service providers to short term
electricity markets within a smart grid context by providing
However, large penetration of EVs can strain the electric prediction of charging event occurrence at the individual
networks, and requires implementation of effective load vehicle level. In particular, this study explores the use of
management strategies, but can also represent a resource to hazard models to estimate the instantaneous charging event
power systems, especially if large shares of intermittent or risk for individual vehicles. The specific objectives of the
fluctuating renewable energy sources exist [2-4], which require present paper are: a) to identify the best time-dependent
increased storage capacity for preservation of grid' s reliability. predictors associated with vehicle use, and b) to test the
The most explored framework allowing managing EV load, residual effect on the hazard of fixed (or slowly varying)
harnessing the distributed storage capacity of grid connected drivers' characteristics.
EVs and integrating EVs into electricity markets is aggregation
[5]. EV aggregators are mediating agents between EV drivers, This analysis will enable informing which instantaneous
electricity market, distribution system operator, and information potentially collected remotely from vehicles is
transmission system operator [6, 7]. To perform their most informative to predict the time to the next recharging
mediating action and participate effectively to electricity event and to demonstrate whether drivers' characteristics could
markets EV aggregators must understand and reliably predict provide additional significant information for the same
charging behaviour of electric vehicles' users. purpose.

Against these background the achievement of an actionable The rest of the paper is organised as follows. Firstly, a
understanding of charging behaviour, which could be methodological section introduces hazard based models with

978-1-5090-1272-5/16/$31.00 2016 IEEE


479
2016 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC)
June 1 -4, 2016, Busan, Korea

time dependent covariates and discusses the issues' in applying where Zij(t) is a covariates' vector which also includes time
them to the problem at hand. I the same session a subsection is varying covariates. Note that in this case the gap times between
devoted to present the datasets used in the study. Then, a charging events are no longer identically distributed and the
section presents and discusses the resulting estimates for the hazard model is not longer strictly speaking a proportional one.
hazard model parameters. Lastly, the paper concludes with a
section summarising the salient points and highlighting the Modelling both unobserved heterogeneity and time
implications of the results. covariates is problematic when time dependent covariates are
not strictly exogenous [1 7]. The strict exogeneity condition as
II. METHODOLOGY expressed by Wooldridge [1 7]. is

A. Analytical Approach Pr[z(t + h)lw t,z(t)] = Pr[z(t + h)lz(t)] \It 0, \lh


In this paper we study the gap times between recurrent >0
charging events. We define these gap times as the time elapsed constant over the interval of observation. This is usually
between two start charging times, so that, in principle, obtained breaking a single gap time into multiple ones in which
modelling this duration would allow predicting when the next the time varying covariate is constant: the generated gap times
charging events would take place. not ending with an event are effectively treated as right
Simplifying, the charging operation process recharging censored. In our analyses we will consider as vehicle use
events could be viewed as a renewal processes, because the variables such cumulative distance, cumulative average speed
charging operation is a sort of regeneration that return an EV and state of charge were constant over the course of a trip (and
into its pristine conditions. Renewal processes have the updated at the start of a trip or at the occurrence of charging
property that gap times Wij tij - tij-1 (between charging any
=
event).
event i - 1 and i for any individual i in population of EV B. Data
drivers) are independent and identically distributed. Under a As part of the Low Carbon London, a project led by
renewal process then, the observation of gap times can be London' s distribution network operator UK Power Networks
analysed as independent durations from different subjects and aiming at investigating the impact of low carbon technologies
modelled with standard survival analysis and for example the on the operation of distribution networks, instrumented Nissan
use of a proportional hazard model would enable incorporating Leaf vehicles were leased to private individuals and companies
the effects of covariates on the hazard function [1 5]. In this and their used was monitored over the course of one year
way one could test whether higher or lower charging events between 201 3 and 201 4. We obtained charging-discharging
hazards (rates of occurrence of charging events) are associated time histories and vehicle use data for twenty of the private
with specific drivers' characteristics. In order to do so a Cox' s vehicles and used the last two months of data for the analyses
proportional hazard model could be estimated of the form in the present paper: March and April 201 4. The reason for
h(wlxij) = ho(w)exp (xij'[3) analysing only the final two months is to use data coming only
from drivers having gained a considerable familiarity with
where ho is an undetermined baseline hazard, xij .is vector using their Leaf in their daily lives.
of covariates associated with the interval before the ith For each monitored vehicle the data available consisted in a
charging event for individual i, and a vector of coefficients. series of events characterised by: event type (driving, charging
In fact the assumption that Wij are independent and or parking), start time ts end time te, State of charge of the
identically distributed is not really tenable, because despite battery at event start SOC(ts)' in %, state of charge of the
conditioning for drivers' characteristics and vehicle usage battery at event end SOC(te)in %, distance travelled during the
metrics, unobserved factors may cause heterogeneity across driving events, d in km, and Average speed during the driving
individuals (i.e. within individual; gap times are correlated). events s in kmlh.
Proportional hazard frailty models can account for this [1 6]. In The original datasets required data pre-processing and
such case the conditional hazard function takes the following cleaning. In particular because of event type classification
form inconsistencies, only the timing, SOC and distance information
hij(wlxij) =
UihO(w)exp (xij'[3) was used to indentify charging events for the analyses
presented here. A charging event was identified for vehicle k
where Ui is an unobserved random effect [1 5]. as occurring when the following condition was satisfied
It is also of interest to consider the effect on the hazard of SOC(te) > SOC(ti) 1\ (d = 0)
the cumulative distance being driven in an inter-charging spell,
the cumulative average speed at which the vehicle is being that is when a positive variation in state of charge occurred
driven, or, the state of charge time path over the spell. These when vehicle was stationary. Positive variations in non
variables are time-varying covariates. Ignoring random stationary events are possible, but they correspond to
heterogeneity, they can be introduced in the hazard model as regenerative charging events, so they are ignored.
follow: From the available data we calculated the time between
charging events for all vehicles over the two months. These are
hij(wIZij(t)) =
ho(w)exp (zij(t)'[3) the gap times used in the hazard based analyses. Likewise we
extracted cumulative distances, cumulative average speeds and

978-1-5090-1272-5/16/$31.00 2016 IEEE


480
2016 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC)
June 1 -4, 2016, Busan, Korea

SOC within these gap times. These are the time dependent distributions of SOC just before charging, distance travelled
explanatory variables whose effect on the instantaneous between charging events and average speed of travel between
charging risk is tested. Fig. 1 shows the distribution of times charging events.
between char in events over our sam Ie as well the

r-

r-
0.5 1.8
r---
r- -
1.6
-
OA
1A -
r- f-- _
1.2

0.3


-
-
0.8 -
0.2
r---
0.6

0.4
0.1

rL
0.2

10 12 14 16 18
"I 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Time elapsed betwenn charging events (gap times) hours SOC at start of charging event

0.Q18
0.06
__ distance ootween charging events(when non zero)

__ average speed(instaces below 10kmlhourare exluded)


0.Q16
__ frt(Burr distribution)
0.05

0.Q14

0.Q12
0.04

0.01

f
0
0.008
f
0
0.03

0.006 0.02

0.004

0.Q1

0.002

140 160 10 20 30 40 50 60 70 80

Distance between charging events km (when non zero) Average speed between charging events kmlhour (intances below 10kmlhour are excluded)

Fig. 1. Distributions of intercharging times and monitored vehicle variables between charing events

From a questionnaire that was administered to the main In the empirical estimation of the hazard model the effects
drivers of the EVs as part of Low carbon London trial a of the following main drivers' attributes were tested: gender,
number of drivers' attributes were available, including: basic age, typical EV use purpose and a range anxiety indicator. Fig.
demographics and information regarding typical EV use 2 shows sununary statistics for these variables.
purposes and well as indicators of their range anxiety level.

978-1-5090-1272-5/16/$31.00 2016 IEEE


481
2016 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC)
June 1 -4, 2016, Busan, Korea

Drivers' demographics Main non-work purpose of


Use of EVto EVUse
16 Commute

'1
14
12
10

6
8
.Male

Female

o +-........
... ---....-
... ....L,
...
4
2
0+---
26-45 46+
Age

Range Anxiety Indicator


Imagine you have to undertake a journey with your EV
along a FAMILIAR ROUTE and a charging facility is
available ONLY AT DESTINATION. What is the
MINIMUM RANGE you would depart with if the journey
were 20 miles long?
15
!!!
Q)
:o
<v
::l 5
z

o
30-35 36-40 41-50 More

Fig. 2. Distribution of driver's characteritics and their stated typical EV use patterns

III. HAZARD FUNCTION ESTIMATION'S RESUSLTS small but significantly different from zero (corr. coefficient -
0.2 P value 0.00). Testing the two specifications enables
We estimate two specifications for the hazard function. understanding whether the two driving variables or the
Specification one uses as time dependent variables the sununary SOC indicator give a better fit. Both specifications
cwnulative driving distance and the cwnulative average speed one and specification two include as fixed (i.e. non time
over the duration of the time interval between charging events. dependent) explanatory variables main drivers characteristics
Specification two uses the SOC value instead. The reason why and typical EV usage patterns as stated by drivers themselves
we use separate specifications is that the SOC level is related to as well as the range anxiety indicator reported in Fig. 2.
both the previous variables. Indeed cumulative distance and
SOC are significantly albeit not so strongly correlated in our Tables 1 and 2 show the estimates of the covariate effects
sample (corr. coefficient -0.6 p value 0.00), whereas the on the hazard function for specifications one and two
correlation of SOC with the cumulative average speed is very respectively.

TABLE!. HAZARD FUNCTION - SPECIFICATION ONE

Variable coef exp(coef) st. error (coef) robust st. error (coef) z stat p value
Cum. distance
0.02 1 1.02 1 0.002 0.004 5.743 9.30E-09
Cum. average speed
0.033 1.034 0.003 0.005 6.5 14 7.30E-ll
Main driver female
0.002 1.002 0. 154 0.296 0.006 1.00E+00
Main driver aged <=45
-0.472 0.624 0. 148 0.3 13 - 1.509 l.30E-01
Range Anxiety Ind.
-0.029 0.97 1 0.005 0.008 -3.685 2.30E-04
Use EV to go to work
-0.492 0.6 12 0.09 1 0.233 -2. 1 15 3.40E-02
Use EV to go to School
-0.449 0.639 0.243 0. 187 -2.393 I.70E-02
N. of observations 9 155 (i.e. measurements of the time dependent variables);N. of charging events 600; N. of vehiclesidrivers 20; Likelihood ratio test=429 on 7 df.

978-1-5090-1272-5/16/$31.00 2016 IEEE


482
2016 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC)
June 1 -4, 2016, Busan, Korea

TABLE IT. HAZARD FUNCTION - SPEClFlCATlON TWO

Variable coef exp(coef) st. error (coef) robust st. error (coef) z stat p value
SOC %
-0.065 0.937 0.003 0.005 - 12. 167 0.000
Main driver female
0.292 1.339 0. 166 0.447 0.654 0.5 10
Main driver aged <=45
- 1.397 0.247 0. 159 0.423 -3.300 0.00 1
Range Anxiety Ind.
-0.044 0.957 0.006 O.ot5 -3.054 0.002
Use EV to go to work
-0.489 0.6 13 0.094 0.3 19 - 1.537 0. 120
Use EV to go to School
-0.474 0.623 0.247 0.300 - 1.58 1 0. 1 10
N. of observations 9 155 (i.e. measurements of the time dependent variables);N. of charging events 600; N. of vehiclesidrivers 20; Likelihood ratio test=886 on 6 df

From specification one' s estimation result we observe that Main drivers aged 45 or younger have an instantaneous
both cumulative distance and average speed have significant charging event rate 75% lower than older drivers;
effects on the hazard function (i.e. the instantaneous charging
event rate). Amongst the fixed effects included in the
An increase by 1 unit of the range anxiety indicator (the
specification above gender and age are not significant. unit of such indicator is miles, see Fig. 2), leads to a 4%
Specification two more than doubles the likelihood ratio decrease in the instantaneous charging rate.
showing that using SOC instead of distance and speed greatly The fust two observations in the list above seem
increases the goodness of fit. However in specification two the reasonable, but the last may appear counterintuitive. Indeed
significance of typical EV use purpose is reduced, while one would expect that having higher battery levels reduces the
drivers 'age become significant. The variability in significance risk of occurrence of changing event. We may also expect that
of the drivers' characteristics in the two specifications is younger people, more prone to errands may charge more
expected given that the sample of drivers is very small (only erratically and less frequently than older individuals that are
20). However it is reassuring to note that the signs of the more likely to have regular patterns enabling them to top-up
effects are stable over the two specifications. their EV battery each night. However, one would expect that
The significant effects (p.value <=0.05) in the two individuals more inclined to rage anxiety would charge more
specifications are more easily interpreted looking at the frequently, not less as these results seem to suggest. One could
exponential of the estimated models' parameters. The speculate that individuals who are more affected by range
exponential of a coefficient represents the hazard ratio for the anxiety tend to use their vehicle less frequently and charge it
corresponding predictor variable. By subtracting one from this only before they use it. However such hypothesis would need
from this ratio one obtains the relative variation of the hazard to be tested further.
give one unit of change of the predictor. Thus the significant In general, the effects of drivers characteristics obtained
effects in specification one can be interpreted as flows: here should be interpreted cautiously because of the sample
An increase by one unit in travelling distance (i.e. an drivers sample at hand. Rather than consider them conclusive,
increase by 1 km) increases the instantaneous risk of a one should interpret them as warnings when developing a
charging event by 2%; predictive model for time-time-to-charging event. Individual
characteristics may play a substantive role and if data regarding
An increase by one unit in travelling speed (i.e. an increase drivers is available in addition to SOC and driving information
by 1 kmlhour) increases the instantaneous risk of a feeds from vehicles, their substantive effect should be tested.
charging event by 3%;
IV. CONCLUSIONS
An increase by 1 unit of the range anxiety indicator (the
This paper, for the fust time to the authors' knowledge, has
unit of such indicator is miles, see Fig. 2), leads to a 3%
used hazard-based duration analysis with time dependent
decrease in the instantaneous charging rate.
covariates to analyse real world charging behaviour with the
Habitual EV Commuters have an instantaneous charging aim to identify possible predictors for the dynamic prediction
event rate 39% lower than the others; of time-to charging events. The motivation for working
towards such prediction models lies in the necessity for
Those drivers using the EV for education purposes have an charging service providers to be able to reliably predict the
instantaneous charging event rate 36% lower than the timing of charging demand for effective participation in
others; electricity markets.
Given that the specification two shows the best fit, we consider The estimated parameters of the hazard model of gap times
it as our main model and we thus interpret comment more in between EV charging events show that the time path of a
depth the corresponding results: vehicle state of charge as well as its cumulative distance and
An increase by one unit in the observed SOC level leads to average speed are strong predictors of the instantaneous rate of
a 6% decrease in the charging event instantaneous risk of a charging events. Drivers characteristics were also found
charging event occurring; affecting the risk of charging events.

978-1-5090-1272-5/16/$31.00 2016 IEEE


483
2016 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC)
June 1 -4, 2016, Busan, Korea

The implication of the findings in this study are that a approach for charging service provider," Transportation Research
Record, vol. Forthcoming,2015.
predictive system for the "time to charging event" should
ideally capture and use instantaneous vehicles parameters (e.g. [ 13] S. Zoepf,D. MacKenzie, D. Keith, and W. Chernicoff, "Charging
Choices and Fuel Displacement in a Large-Scale Demonstration of
sate of charge and cumulative driving distances) but should not Plug-In Hybrid Electric Vehicles," Transportation Research
neglect other "modulating factors", such as drivers' Record: Journal of the Transportation Research Board, vol. 2385,
demographics or range anxiety. In order to test how substantive pp. 1- 10, 12/0 11 2013.
are such modulating factors in prediction, larger drivers' [ 14] N. Daina,A. Sivakumar,and 1. Polak, "Patent and latent predictors
sample need to be adopted. As EVs adoption rates are steadily of electric vehicle charging behaviour," Transportation Research
Record, vol. Forthcoming,2015.
increasing, such data will be increasingly easier to obtain
without need of ad hoc electric vehicle trials. As next step of [ 15] R. J. a. Cook, "The Statistical Analysis of Recurrent Events," in
Statistics for Biology and Health, F. L. a. Jerald and SpringerLink,
the present work we intend to investigate how substantively
Eds.,ed,2007.
important individual characteristics are in prediction and thus
[ 16] T. M. Therneau, P. M. Grambsch, and V. S. Pankratz, "Penalized
whether it is worth investing effort to collect the associated Survival Models and Frailty," Journal of Computational and
data in an operational implementation. Graphical Statistics, vol. 12,pp. 156- 175,2003.

[ 17] 1. M. Wooldridge, Econometric analysis of cross section and panel


ACKNOWLEDGMENT data, 2nd ed. Cambridge,Mass.: MIT Press,2010.

We thank UK Engineering and Physical Sciences Research


Council for supporting this work under awards EP/LOO1 03911
and EP/I03883711 and UK Power Networks for providing the
EV charging datasets used for the analysis.

REFERENCES

[ 1] lEA, "Technology Roadmaps. Electric and plug-in hybrid electric


vehicles. Updated 20 1 1.," International Energy Agency20 1 1.
[2] W. Kempton and S. E. Letendre, "Electric vehicles as a new power
source for electric utilities," Transportation Research Part D:
Transport and Environment, vol. 2,pp. 157-175, 1997.

[3] W. Kempton and J. Tomic, "Vehicle-to-grid power


implementation: From stabilizing the grid to supporting large-scale
renewable energy," Journal of Power Sources, vol. 144, pp. 280-
294,2005.
[4] C. Quinn, D. Zimmerle, and T. H. Bradley, "The effect of
communication architecture on the availability, reliability, and
economics of plug-in hybrid electric vehicle-to-grid ancillary
services," Journal of Power Sources, vol. 195, pp. 1500-1509,
2010.
[5] R. 1. Bessa and M. A. Matos, "Economic and technical
management of an aggregation agent for electric vehicles: a
literature survey," European Transactions on Electrical Power,
vol. 22,pp. 334-350,20 12.
[6] R. J. Bessa and M. A. Matos, "Global against divided optimization
for the participation of an EV aggregator in the day-ahead
electricity market. Part I: Theory," Electric Power Systems
Research, 2012.

[7] R. J. Bessa and M. A. Matos, "Global against divided optimization


for the participation of an EV aggregator in the day-ahead
electricity market. Part 11: Numerical analysis," Electric Power
Systems Research, 20 12.

[8] A. P. Robinson, P. T. Blythe, M. C. Bell, Y. HUbner, and G. A.


Hill, "Analysis of electric vehicle driver recharging demand
profiles and subsequent impacts on the carbon content of electric
vehicle trips," Energy Policy, vol. 6 1,pp. 337-348, 101/ 2013.
[9] T. Franke and 1. F. Krems, "Interacting with limited mobility
resources: Psychological range levels in electric vehicle use,"
Transportation Research Part A: Policy and Practice, vol. 48, pp.
109-122,2013.
[ 10] T. Franke and J. F. Krems, "Understanding charging behaviour of
electric vehicle users," Transportation Research Part F: Traf
f ic
Psychology and Behaviour, vol. 2 1,pp. 75-89, 1 1112013.

[ 1 1] N. Daina, "Modelling electric vehicle use and charging behaviour,"


PhD thesis,Civil and environmental engineering,Imperial College
London,London,2014.
[ 12] c. Latinopoulos, 1. Polak, and A. Sivakumar, "Joint charging and
parking choices of electric vehicle drivers: Decentralized control

978-1-5090-1272-5/16/$31.00 2016 IEEE


484

You might also like