A data-driven approach for characterising the charging demand of electric vehicles:A UK case
study
Erotokritos Xydas a ,⇑,Charalampos Marmaras a ,Liana M.Cipcigan a ,Nick Jenkins a ,Steve Carroll b ,Myles Barker b
a Cardiff University,School of Engineering,The Queen’s Buildings,The Parade,CF243AA Cardiff,Wales,
UK
b
Cenex,Innovation Centre,Loughborough University Science &Enterprise Parks,Oakwood Drive,LE113QF Loughborough,UK
h i g h l i g h t s
21,918charging events from 255different charging stations in UK were analysed. A data pre-processing methodology for dealing with EVs charging data was presented. A data mining model was developed to analyse the EVs charging data.
A fuzzy logic decision model was developed to characterise the EVs charging demand.
a r t i c l e i n f o Article history:
Received 4August 2015
Received in revised form 21October 2015Accepted 22October 2015
Keywords:
Characterisation model Data mining Data analysis
Electric vehicles charging events
a b s t r a c t
As the number of electric vehicles increases,the impact of their charging on distribution networks is being investigated using different load profiles.Due to the lack of real charging data,the majority of these load impact studies are making assumptions for the electric vehicle charging demand profiles.In this paper a two-step modelling framework was developed to extract the useful information hidden in real EVs charging event data.Real EVs charging demand data were obtained from Plugged-in Midlands (PiM)project,one of the eight ‘Plugged-in Places’projects supported by the UK Office for Low Emission Vehicles (OLEV).A data mining model was developed to investigate the characteristics of elec-tric vehicle charging demand in a geographical area.A Fuzzy-Based model aggregates these characteris-tics and estimates the potential relative risk level of EVs charging demand among different geographical areas independently to their actual corresponding distribution networks.A case study with real charging and weather data from three counties in UK is presented to demonstrate the modelling fr
amework.Ó2015The Authors.Published by Elsevier Ltd.This is an open access article under the CC BY license (
1.Introduction
Electric Vehicles (EVs)offer reduced transportation related emissions,reduce the energy cost of driving and in some cases eliminate the use of fossil fuels.The total electricity demand is expected to grow as the number of EVs increases [1].The impact of EVs charging on distribution networks has been investigated in the literature.The majority of these studies use synthetic data to assess the impact of the EVs charging load due to limited access to real EVs charging data.In [2–19]data from travel surveys are used to create EVs charging load profiles,assuming that EVs are travelling like conventional internal combustion engine vehicles.
Although EVs adoption is at an early stage,some utilities and aggregators are already collecting information from charging sta-tions.A limited number of EVs pilots exist around the world,allow-ing some preliminary studies on charging demand profiles.In [20],statistical analysis of 4933charge events in the Victorian EVs Trial in Australia was performed.Statistical models for charge duration,daily charge
frequency,energy consumed,start time of charge event,and time to next charge event were estimated to express the uncertainty of usage patterns due to different user behaviours.Data from the Western Australian Electric Vehicle Trial (2010–2012)were analysed in [21,22],investigating the drivers’recharging behaviours and patterns.In [23],7704electric vehicle recharging event data from the SwitchEV trials in the north east of England were used to analyse the recharging patterns of 65EVs.The results showed that minimal recharging occurred during off peak times.In [24]data from the same project were combined
/10.1016/j.apenergy.2015.10.151
0306-2619/Ó2015The Authors.Published by Elsevier Ltd.
This is an open access article under the CC BY license (/licenses/by/4.0/).
⇑Corresponding author.
E-mail address:xydase@gmail (E.Xydas).
with low voltage smart meter data from Customer Led Network Revolution(CLNR)project and the impact of the combined demand profile was assessed on three different distribution networks.The resul
ts showed that the spatial and temporal diversity of EVs charging demand reduce its impact on those distribution networks. Finally,data from over580,000charging sessions and from2000 non-residential electric vehicle supply equipment’s(EVSE)located in Northern California were analysed in[25].The scope of this analysis was to investigate the potential benefits of smart charging utilising the extracted information regarding the actual trips and customer characteristics.
Monitoring the charging events will inevitably create large volumes of data.These data require effective data mining methods for their analysis in order to extract useful information.In[26–28] various data mining techniques were utilized to address challenges in the energy sector,such as load forecasting and profiling.In [29–31]data mining modelling frameworks were applied to electricity consumption data to support the characterisation of end-user demand profiles.
In this paper,a framework was developed to characterize the EVs charging demand of a geographical area.The technical contri-butions of this paper are summarised below:
(i)Real EVs charging data from UK were acquired and analysed.
The diverse data were organised and classified into attri-butes.To the authors’best of knowledge,this is thefirst time that real EVs charging data are presented using this level of detail.
(ii)A comprehensive data cleaning and formatting methodology is presented,developed specifically for dealing with EVs charging data.
(iii)A data mining model was developed to extract the useful information.Three key characteristics of EVs charging demand in a geographical area were investigated using the proposed methodology,namely shape of the typical daily profile,predictability with respect to weather and trend.
Clustering,correlation and regression analysis were per-formed to study each characteristic,using factors to quantify them.Analysing these characteristics resulted in assessing the potential risks and uncertainties which affect the mid-term normal operation of the corresponding distribution network.
(iv)A fuzzy logic decision model was developed that aggregates the three factors into one‘‘risk level”index.The‘‘risk level”
index was defined in order to characterize the EVs charging demand,reflecting its potential impact on the energy demand in a geographical area.Areas with high‘‘risk level”
values imply a potential risk for the mid-term normal operation of the distribution networks and such analysis could be important for the distribution network operator (DNO).No similar research work that q
uantifies the mid-term relative risk of the EVs charging demand among different geographical areas independently to their actual corresponding distribution networks was done so far.
(v)Furthermore,this paperfills a gap in the literature related to handling real EVs charging data,by proposing a complete data analysis methodology.
The rest of the paper is organized as follows:Section2describes the real EVs charging data analysed.In Section3the proposed methodology to characterize the EVs charging demand is illus-trated.A case study is presented in Section4,applying the model on real EVs charging events from UK to study the charging demand characteristics,and assess their potential impact.Finally,conclu-sions are drawn in Section5.2.Data description
EVs charging demand data were obtained from the Plugged-in Midlands(PiM)project(uk/). The Plugged-in Midlands project,managed by Cenex,is one of the eight‘Plugged-in Places’projects supported by OLEV,the Office for Low Emission Vehicles in the UK.Two datasets were provided by Cenex,with information regarding the charging events and charging stations respectively.The charging events dataset con-sists of21,918charging events from255different charging stations and587unique EVs drivers.The charg
ing event dataset includes information about the connection/disconnection times and the energy of each charging event for the period of2012–2013with event-occurrence granularity.The charging station dataset con-tains time-independent information regarding the location and technical specifications of all charging he charging power rate).The contents of the two datasets are listed in Tables 1and2.
An additional dataset was acquired from the UK Met Office, with information regarding the weather in the Midlands,the geo-graphical area under study.This dataset includes the values of var-ious weather air temperature)with daily granularity for the period of2012–2013.The weather attributes are listed in Table3.
3.Methodology
The characterisation framework consists of three models:(i) Data Pre-processing Model,(ii)Data Mining Model and(iii)Fuzzy Based Characterisation Model.The Data Pre-processing Model pro-vides data merging,cleaning and formatting to prepare the data
Table1
Charging event data.
Attribute name Attribute description
Connection time Start time of charging event
in dd/mm/yyyy hh:mm format Disconnection time End time of charging event in
dd/mm/yyyy hh:mm format Energy drawn Energy demand of charging event in kW h User Unique ID for every EV1,EV2etc.
Charging station Unique ID for every charging station
Table2
Charging station data.
Attribute name Attribute description
Charging station Unique ID for every charging station
Latitude Latitude of charging station’s location
Longitude Longitude of charging station’s location
Road The road name of charging station’s location Post code The post code of charging station’s location
County The county name of charging station’s location Location category e.g.Private Parking,Public Parking etc.
Location subcategory e.g.Public Car Park,Public On-street etc.
Ownership e.g.Dealership,Hotel,Train Station
Host Name of the charging station host
NCR Whether or not the charging station is
registered on the
National Charging Registry(NCR)of UK Manufacturer The charging station manufacturer
Supplier The operator of charging station
Charger type Power rate of charging station in kW
Connector1Socket Pin 3Pin,5Pin etc.
Connector2If exists,the second Socket Pin Type
Mounting type e.g.Ground,Wall,Wall(tethered)
764 E.Xydas et al./Applied Energy162(2016)763–771
for the Data Mining model.The Data Mining Model consists of three modules namely Clustering Module,Correlation Module and Regression Module.These modules were used to investigate the shape of the typical daily profile,the predictability with respect to weather and the trend of EVs charging demand respectively. The Fuzzy Based Characterisation Model aggregates the outputs of the Data Mining model into a‘‘risk level”index of EVs charging demand in a geographical area using fuzzy logic.The characterisa-tion framework is illustrated with Fig.1.
3.1.Data pre-processing model
Data of the Connection Time,Disconnection Time,Energy Corrupted or missing data are not a rare phenomenon in such com-plex communication networks.However,a careful analysis at this stage is also beneficial tofind the location or the station’s ID from where the corrupted data are recorded,an indi
cation of an abnor-mal operation.
The next stage of the Data Pre-processing model is the Forma-tion stage.The EV dataset was formatted using a Matlab script into three time series;an hourly power time series,a daily peak power time series and a monthly energy time series.The hourly power time series was transformed into daily vectors(each of24values) and forwarded to the Clustering Module,whereas the monthly energy time series was forwarded to the Regression Module.All the data attributes of the Weather dataset were formatted into daily time series and merged with the daily peak power time series.The resulting(combined)time series was forwarded to the Correlation Module.The data pre-processing procedure is presented in Fig.2.
3.2.Data mining model
The Data Mining Model consists of a Clustering Module,Corre-lation Module and Regression Module.These modules were used to investigate the shape of the typical daily EVs charging demand pro-file,the predictability with respect to weather and the trend of EVs charging demand respectively.
3.2.1.Clustering module
The clustering module creates typical daily EVs charging
Table3
Weather data.
Attribute name Attribute description
Max air temperature Daily maximum air temperature(°C)
Min air temperature Daily minimum air temperature(°C)
Mean air temperature Daily average air temperature(°C)
Mean wind speed Daily average wind speed(knots)
Max gust Daily maximum wind speed(knots)
Rainfall Daily precipitation(mm)
Daily global radiation Daily amount of solar energy falling on a horizontal
surface(kJ/m2)
Daily sunny hours Daily sunshine duration(h)
Fig.2.Data pre-processing model.
E.Xydas et al./Applied Energy162(2016)763–771765
according to its distance with the nearest cluster centroid.Then,the new cluster centroids are obtained from the average of the daily vec-tors for the corresponding cluster.This process is repeated until the distances between the daily vectors and the corresponding cluster centroids are minimized.This is explained mathematically by Eq.(1):
min
c
X k i ¼1X x 2c i
k x Àl i k 2
ð1Þ
where c i is the set of daily vectors that belong to i th cluster,x expresses the corresponding daily vector in c i and l i is the position of the i th cluster centroid.
The method requires the number k of clusters to be defined a priori.The Davies–Bouldin evaluation criterion was used to calcu-late the number k of clusters [34,35].This criterion is based on a ratio of within-cluster and between-cluster distances and is defined by Eq.(2):
DB ¼1k X k i ¼1max
j –i d i þ d j
d ij
!ð2Þ
where d i
is the average distance between each point in i th cluster and the centroid of i th cluster. d
j is the average distance between each point in i th cluster and the centroid of j th cluster.d ij is the dis-t
ance between the centroids of i th and j th clusters.The maximum value of this ratio represents the worst-case within-to-between cluster ratio for i th cluster.The ‘‘best”clustering solution has the smallest Davies–Bouldin index value.Therefore,an additional step exists to evaluate the centroid selection for our dataset.A range of 1–20clusters was considered,where 20was found to be a rea-sonable maximum value [36],and the best number of clusters within this interval was calculated using an iterative process.By applying the k-means clustering method to the dataset,the k cluster centroids c i are obtained,along with the number of vectors w i assigned to each cluster.The followed steps of the Clustering Mod-ule are presented in Fig.3.
The most representative cluster centroid (highest value of w i )was used to create the typical daily EVs charging demand profile k ¼
E peak
E total
Á100%ð3Þ
where E peak is the charging load during the peak hours and E total is the total daily charging load.
3.2.2.Correlation module
According to [38],weather affects road traffic congestion and the driving behaviour of car owners.In [39–41],the factors which affect the fuel consumption of EVs were analysed.Cold weather decreases the efficiency of the batteries performance.Additionally,heating the interior of EVs drains significantly the battery.In [42],the impact of cold ambient temperatures on running fuel use was investigated.Considering EVs on the roads,the weather will also affect their energy consumption and thus their charging demand.Identifying hidden strong relationships between weather attributes and load demand improves the forecasting accuracy of a prediction model [43].
The Pearson’s Correlation Coefficient (r )was used in this module to measure the correlation between the weather attribute values and the daily peak power of EVs charging demand in a geo-graphical area.The maximum absolute correlation coefficient value of all peak power-weather pairs identifies the most influen-tial weather attribute.
3.2.3.Regression module
The scope of this module is to investigate the monthly change of the EVs charging demand.A Growth Ratio (GR)index was defined as the ratio between the growth rate of EVs charging demand and the av
erage monthly EVs charging demand.Linear regression anal-ysis was applied on the EVs charging demand time series,in order to calculate the mathematical formula describing the relationship between monthly EVs charging demand (Y in kW h)and time (X in months).The formula is described with Eq.(4):
Y ¼b 0þb 1X þe ;
characteriseð4Þ
where b 0and b 1are the constant regression coefficients and e is the random disturbance (error).
The slope b 1expresses the monthly growth rate of EVs charging demand (in kW h/month).The constant regression coefficients were calculated using the Least Squares Method described in [44].Having b 1,the GR index is calculated with Eq.(5).
GR b 1month
100%;ð5Þ
where month is the average monthly EVs charging demand.3.3.Fuzzy based characterisation model
The goal of this model was to characterise the EVs charging demand of a geographical area according to the information about the shape of the typical daily profile (k index),the predictability with respect to weather (r )and the trend of EVs charging demand (GR index).To this end,a ‘‘risk level”index was defined to aggre-gate the potential underlying risks from these characteristics.A fuzzy-logic model was developed to capture the fuzziness of these risks and calculate the ‘‘risk level”index.Fuzzy Logic Models are useful for risk assessment purposes under such conditions [45].The Fuzzy Based Characterisation Model is illustrated with Fig.4.The validity of the risk characterisation model is based on the following considerations/assumptions:
i.The magnitude and duration of the peak of the typical EVs charging demand profile (captured by k index)are underly-ing risk factors for the distribution network,as they affect the transformer/circuit loading and the voltage profile.
Fig.3.Clustering Module flowchart.
766 E.Xydas et al./Applied Energy 162(2016)763–771
ii.The change over time of EVs charging demand (described with GR index)affects the long term decision regarding the planning of the network reinforcement.The aggressive-ness of EVs charging de
mand change over time in a geo-graphical area is also a potential risk for the network’s operation.
iii.The predictability of EVs charging demand with respect to
weather in a geographical area (captured by r ),affects the accuracy of a forecasting model.Decisions taken based on a forecast are subject to the forecasting accuracy,indicating a risk for the decision maker.
iv.Analysing the EVs charging demand characteristics in a geo-graphical area results in assessing the risks and uncertainties which will affect the mid-term normal operation of the dis-tribution network of the corresponding geographical area.v.As an electric power network model was not used to analyse the related actual charging demand characteristics,this study quantifies only the relative risk between different geographical areas.The ‘‘risk level index”is not defined in absolute terms and thus it is used to classify relatively the level of these risks (due to EVs charging)among different geographical areas independently to their actual corre-sponding distribution networks.The linguistic values used to express the input variables are Low (L ),Medium (M )and High (H ).Triangular membership functions are used to calculate the Degree-Of-Membership (DOM)for each of them,as shown in Figs.5–7.In contrast to other kind of mem-bership functions (e.g.Trapezoids),triangular membership func-tions are very sensitive to changes of the variables and thus this increase the accuracy.
The output is fuzzified into nine fuzzy regions represented by linguistic variables;very very high (VVH),very high (VH),high (H ),medium high (MH),medium (M),medium low (ML),low (L),very low (VL)and very very low (VVL),as shown in Fig.8.The rule table is given in Table 4.
The design of the rule table is based on the assumption that each of the input indicators affect equally the ‘‘risk level”index.According to the best of the authors’knowledge,there is no research work that quantifies the level of influence of the related indicators (k index,r and GR)to the operation of an electricity dis-tribution network.A further investigation is necessary to under-stand the relative impacts of these variables on the normal operation of an electricity distribution network,but this is out of the scope of this paper.
The Mamdani type inference was used (also known as the max–min inference method),which utilizes the minimum function for the implication of the rules.Defuzzification was performed using the centre of gravity (CoG)method [46–48].This method
finds
Fig.4.Fuzzy Based Characterisation Model.
Table 4Rule table.r
k
GR L
M H H
L VVL VL L M VL L ML H L ML M M
L L ML M M ML M MH H M MH H L
L M MH H M MH H VH H
H
VH
VVH
162(2016)763–771767
版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。
发表评论