Content uploaded by Hung Cao
Author content
All content in this area was uploaded by Hung Cao on Aug 28, 2020
Content may be subject to copyright.
Discovering EV Recharging Patterns through an
Automated Analytical Workflow
Ren´
e Richard†∗ , Hung Cao∗, Monica Wachowicz∗
∗People in Motion Lab, University of New Brunswick, Canada
†National Research Council of Canada, Digital Technologies Research Centre, Fredericton, Canada
{rene.richard, hcao3, monicaw}@unb.ca
Abstract—The vision for smart cities is to provide a core
infrastructure that enables a good quality of life for their citizens
and the sustainable management of natural resources. Towards
this vision, supporting the adoption of Electric Vehicles (EV)
contributes to improved air quality, sustainable mobility, and util-
ity distribution. Fostering EV adoption contends with concerns
typically centered on vehicle range and costs. An understanding
of EV charging patterns is therefore crucial for optimizing
charging infrastructure placement and managing operational
costs. Towards this end, this paper proposes an automated
analytical workflow to gain insight from a large volume of real
operational data from EV charging stations. The research goal
is to establish a mechanism to descriptively analyse the EV
charging data and to thoroughly diagnose whether low-demand
charging station groupings can effectively be identified using
spatio-temporal features and hierarchical clustering. Preliminary
results suggest agglomerative clustering is effective at grouping
similar charging stations together when considering spatial and
temporal features of recharge events.
Index Terms—agglomerative hierarchical clustering, EV adop-
tion, charging infrastructure patterns, automated machine learn-
ing flow
I. INTRODUCTION
Globally, national and local government commitments to
electrify the transport sector will have an impact on smart
cities. In Canada, 24% of GHG emissions come from cars.
New Brunswick has the third highest per-capita GHG emis-
sions in Canada and thirty percent of these emissions are
attributable to the transportation sector [1]. By decoupling
vehicles from the immediate consumption of fossil fuels,
options for supporting mobility from renewable resources
increase.
A significant increase in EV adoption will require adequate
public charging infrastructure. One commonly stated barrier to
widespread EV adoption is driver range anxiety [1], [2]. While
increasing the availability of public charging infrastructure can
appease some concerns, charging infrastructure operators do
not want to invest in charging stations because they tend to
be less profitable in early EV adoption contexts. Additionally,
changing population demographics in some regions can pose
further challenges to EV uptake. Enthusiasm around vehicle
electrification has been associated with a youthful user base
and financial incentives, which some regions struggle to pro-
vide [3].
Broader EV adoption will require a public recharging net-
work to serve a population with different living and park-
ing situations (e.g. multi-tenant dwellings) [4]. However, an
understanding of EV usage patterns is also crucial to foster
adoption while managing costs and optimizing placement of
charging infrastructure. Although there exists a rich literature
on different algorithms developed to find EV usage patterns,
lacking is the use of real-world EV charging events from
public infrastructure. Analyses are often based on assumptions
that EV traveling profiles are similar to that of conventional
internal combustion engine vehicles.
By using real-world EV charging event data, related demand
characteristics can provide a better assessment of increased EV
adoption and the potential impact on energy demand. The main
challenge is to develop an analytical workflow that allows the
creation of a containerized assembly, which fully automates
analytical tasks and facilitates the reproducibility, sharing and
distribution of the computational environment. Towards this
end, this paper proposes an end-to-end analytical workflow to
facilitate charging pattern analysis from an early EV adoption
context in Atlantic Canada. In total, nine automated tasks are
developed and a case study using real-world charging event
data from station operators in Atlantic Canada is used for the
implementation.
The main contributions of this paper are:
•Automated analytical tasks for discovering EV charg-
ing patterns are still nonexistent despite the need for
analytical workflows capable of handling the expected
volume and complexity of real-world EV charging events.
The pioneering research work on the analytical workflow
proposed in this paper is a first step towards this direction.
•The empirical results of this research work contribute to
advance scientific knowledge on the current EV adop-
tion patterns that play a role in planning and extending
charging infrastructure design for electric vehicles.
The rest of the paper is organized as follows. In Section II,
we explore related works. Section III describes the nine auto-
mated tasks developed for our proposed analytical workflow
and Section IV provides a detailed description of real-world
EV charging event data and computational environment used
for implementation. In Section V, we discuss the results.
Finally, Section VI concludes and indicates future research
work.
II. RE LATE D WOR K
A significant proportion of the work studying EV charging
patterns uses data from sources other than real-world EV
charging events in order to assess the impact of broad EV
adoption on distribution networks. This is apparent in a recent
review paper by Hardmana et al. [5] on consumer preferences
with regards to EV charging infrastructure, which lists studies
that employed surveys, interviews, modeling and vehicle GPS
data in addition to a small number of studies using EV
charging equipment information. By using real-world charging
event data, related demand characteristics can provide a better
assessment of increased EV adoption and the potential impact
on energy demand.
A more complete picture of EV infrastructure usage patterns
can be formulated by combining data from different sources.
However, scarcity in complimentary data sources is also a con-
cern. Ashkrof et al. [2] have studied EV user travel behaviors
in the Netherlands and point out the main limitations of their
work was related to the low number of Battery Electric Vehicle
(BEV) driver participants. In order to compensate for this,
Hybrid Electric Vehicle (HEV) and Plug-in Hybrid Electric
Vehicle (PHEV) drivers were added to the study. Despite the
limitation in targeting BEV owners uniquely, the authors found
that route attributes such as travel time, charging infrastructure
characteristics en route to and at the destination of travel in
addition to charging wait times and State-of-Charge (SOC)
influences route selection and charging behavior considerably.
In their study, the authors note that SOC and lack of charging
opportunities are the main concerns of EV drivers.
The classification and modeling of energy usage behavior
is core to improving emerging applications and services.
Enhancements in services allow for smoothing of frequent
peaks and imbalances. In the energy domain, clustering is
commonly used to group similar consumers, predict future
energy demand or detect outliers.
Our use of real operational data from EV charging stations
advances our understanding of EV charging behavior. To the
best of our knowledge, no other work has implemented an EV
charging analysis workflow that facilitates automated, parallel,
spatio-temporal analysis in a containerized environment. This
work aims to answer the question of whether agglomerative
clustering can effectively identify low-energy demand clusters
by grouping under-utilized (and normally utilized) charging
stations together using spatio-temporal features.
III. METHODOLOGY
The overall goal of building an analytical workflow is
to explore patterns that develop around EV charging infras-
tructure usage. Therefore, the proposed workflow supports
automated tasks that are essential in providing new insights
on whether low-demand charging stations can effectively be
identified to ultimately assess their impact on EV adoption
(i.e. Descriptive analytical results). These automated tasks are
also important to diagnose the cause of phenomena observed
at charging stations based on clustering of real-world charging
events (Diagnostic analytical results). Figure 1 illustrates the
data flow and processing tasks of the proposed end-to-end
analytical workflow which is designed to operate in an fully
automated manner and facilitates the reproducibility, sharing,
and distribution of the computational environment.
In total, nine automated analytical tasks are designed to
analyse a massive amount of real-world EV charging data in a
fast parallel processing environment using different techniques
to track, manage, and compare the analytical results.
Each analytical task can be described as one of the follow-
ing:
•Data Cleaning Task: This task receives the raw data
input from the public charging stations. The raw data
can be defined as a set of files R={R1, R2, ..., Rn}
where each file Ri3[r1, ..., rm]data rows in which
each data row rjowns kattributes. The goal of this data
cleaning task is to clean each raw data file in the set R. A
core function is developed to guarantee the data quality
and produce a set of cleaned files C={C1, C2, ..., Cn}
by eliminating errors, inconsistencies, duplicated and
redundant data rows, and handling missing data.
•Data Integration Task: This task is known as a practice
of consolidating data from various data files into a single
dataset. A variety of files from the cleaned set Cin the
previous task can be used as the input for this task. The
output of this task is a unique file Ithat merged all
attributes from set Cinto one big table.
•Data Fusion Task: Different from the data integration
task, a data fusion task usually involves combining multi-
ple data sources followed by reduction or replacement for
the purpose of better inference. In our proposed analytical
workflow, multiple integrated data files Ican be pushed
into the data fusion task and produce more consistent,
accurate, and useful data files Fthat serve a more narrow
set of application workloads.
•Data Contextualization Task: The aim of this task is to
enrich the fused data files Fstep by step by adding new
attributes to each data row according to a specific context.
This task is defined by a contextualization function that
can produce a set of new data rows Pi∈Pusing
contextualization parameters hΨ1,Ψ2, ...ito add new
attributes to the fused data rows Fi∈F.
∀Fi∈(F1, F2, . . . , Fn) : Fi= (f1, ..., fm)
F= (F1, . . . , Fn)hΨ1,Ψ2,...i
−−−−−−−→ P= (P1, . . . , Pn)
∀Pi∈(P1, . . . , Pn) : Pi= (p1, ..., pm, Context1, C ontext2, . . . )
(1)
It is crucial in transforming fused data rows generated by
EV charging events into semantically enriched data that
are needed as an input to the next analytical tasks.
•Data Descriptive Task: To gain an overall understanding
from the contextualized data in the previous task, this
task performs several descriptive statistical functions.
They include frequency measurement, central tendency
measurement, dispersion or variation measurement, and
position measurement.
Data
Cleaning
Data
Cleaning
Data
Transforma
tion
Data
Transforma
tion
Data FusionData Fusion
Data
Extraction
Data
Extraction
Data
Integration
Data
Integration
Data
Contextuali
zation
Data
Contextuali
zation
Data
Aggregation
Data
Aggregation
Data
Diagnostic
Data
Diagnostic
Analytical Workflow
Cleaned
Data
Integrated
Data
Fused
Data
Contextualized
Data
Extracted
Data
Transformed
Data
Aggregated
Data
Diagnostic
Analytical
Results
Raw Data
Data
Descriptive
Data
Descriptive
Descriptive
Analytical
Results
(Spatio-Temporal
Clustering Patterns)
Charging Level 3
Charging Level 2
Charging Level 1
Cable connect to charger
Control &
Communication
Control &
Communication
Control &
Communication
AC
AC
DC
Build-in cable protection
Smart recharge
stations
AC
Fig. 1: The proposed automated end-to-end analytical workflow.
•Data Extraction Task: This task is defined by an ex-
traction function that can produce a subset of data rows
Ethat are extracted from a set of contextualized data
rows Pusing extraction (filtering) parameters hΩ1,Ω2, ...i
executed on a selected attribute (or a set of selected
attributes) of a set of contextualized data rows P.
∀Pi∈(P1, . . . , Pn) : Pi= (p1, ..., pm, Context1, C ontext2, . . . )
P= (P1, . . . , Pn)hΩ1,Ω2,...i
−−−−−−−−−→
on attributes (ei)
E= (E1, . . . , En)
∀Ei∈(E1, . . . , En) : Ei= (att1, att2, . . . ),∀att ⊂ {e1, . . . , em}
(2)
•Data Transformation Task: This task is defined by
a transformation function that performs transformation
operations using parameters hΥ1,Υ2, ...iexecuted on a
selected attribute (or a set of selected attributes) of a set
of extracted data rows Eor aggregated data rows Ato
produce a set of new data rows T.
(E1, . . . , En) : Ei= (att1, att2, . . . )
E= (E1, . . . , En)∨A= (A1, . . . , An)hΥ1,Υ2,...i
−−−−−−−→ T= (T1, . . . , Tn)
∀Ti∈(T1, . . . , Tn) : Ti= (T rans value1, T r ans value2, . . . )
(3)
•Data Aggregation Task: Aggregation is a mathematical
operation (e.g. sum, average, count, minimum) that takes
multiple attributes of many rows and returns a single
value. This task is defined by aggregation parameters
hΦ1,Φ1, ...iexecuted on a selected attribute (or a set of
selected attributes) of a set of transformed data rows T
to produce a set of new data rows A.
∀Ti∈(T1, . . . , Tn) : Ti= (T rans value1, T rans value2, . . . )
T= (T1, . . . , Tn)hΦ1,Φ1,...i
−−−−−−−−−−−−−−−→
on attribute Tr ans v aluei
A= (A1, . . . , Am)
∀Aj∈(A1, . . . , Am) : Aj= (Agg value1, Ag g value2, . . . )
(4)
•Data Diagnostic Task: The aim of this task is to find
the patterns from transformed data using a hierarchical
agglomerative clustering algorithm [6]. The raw recharge
events are transformed into a per-minute kWh energy
delivery format for each station recharge event, which
also includes station latitude and longitude coordinates.
Then, the PCA technique is used to reduce the dimensions
of the transformed data, since the number of spatial and
temporal features is very high. Finally, an agglomerative
clustering model is utilized to fit this data. The clustering
was performed on weekly, monthly and seasonal data
partitions to provide results for different time windows.
A priori unknown schemes inherit in the charging data
were identified with this unsupervised learning approach;
grouping stations in terms of their similarity.
IV. IMPLEMENTATION
A. Data Collection
EV charging opportunities are often grouped in three levels
based on voltage, current and typical charging times. These
levels are : Level 1 (L1), Level 2 (L2) and Level 3 (or
DC Fast) [4]. Our study used real operational data from
public electric vehicle charging stations provided by the New
Brunswick Power Corporation. For this work, we selected
the EV charging events that occurred between the dates of
April 2019 and April 2020 at 37 Level-2 (L2) and 26 Level-
3 (L3) public charging stations. The total number of charging
events included in the analysis was 9,505. The total number of
minutes spent recharging on L2 and L3 charging infrastructure
was 551,635 minutes, which rounds up to 9,194 hours. The
total amount of energy transferred to vehicles during the study
period was 97,148.65 kWh.
Table I describes the features included in the raw EV
charging data set. The raw data were fused with charging
station location information and transformations were applied
in order to feed the downstream processing.
TABLE I: Raw Data
Column Name Description
Connection ID Unique identifier for a connection
Recharge start time (local) Timestamp denoting start of charging
event
Recharge end time (local) Timestamp denoting end of charging event
Account name Unused (all null)
Card identifier Unique identifier for a charging
plan member
Recharge duration
(hours:minutes)
Duration of charge event
Connector used Connection used during charge event
Start state of charge (%) State of charge percentage at beginning of
charging event
End state of charge (%) State of charge percentage after charging
event is complete
End reason Charge event end reason
Total amount Unused (all null)
Currency Unused (all null)
Total kWh Energy transferred to vehicle during
charging event
Station Unique identifier for a charging station
B. Computational Environment
The software programs used in this work were packaged
using Docker [7] containers in order to ensure a reproducible
computational environment and to facilitate the distribution
or extension of experimental workflows. A local Spark [8]
cluster with 18 worker threads and 20GB of RAM was
used to process the data. The data processing and analytical
workflow was implemented using custom-written code, which
used a standard scientific Python stack comprised of PySpark,
Pandas, scikit-learn, NumPy and SciPy.
C. Analytical Workflow Implementation
Fig. 2 highlights noteworthy aspects of the analytical work-
flow implementation. We leveraged MLflow [9] to manage the
development life-cycle of the proposed automated analytical
workflow outlined in Fig. 1. The numbered boxes in Fig. 2
represent individual Spark jobs. The data flow is such that the
output of one job is the input for the next job. Input and output
file names contain parameter values that were used when
calling the workflow’s scripts. The grey elements represent a
job’s input file(s). The blue elements represent a job’s output
file(s).
Program elements were executed in sequence using Shell
commands that called parameterized Python scripts. Examples
of Bash scripts for executing workflow tasks 1 and 5 can be
found here1. What follows is a description of each data pro-
cessing workflow element from loading the initial raw data to
applying the clustering algorithm. The detailed implementation
of each task in the workflow are described as follows:
•Task (1): We developed the one way hash.py script
to import raw event data and cast column elements to
appropriate types. Additionally, a one-way hash function
was applied to the Card identifier column and the output
was saved to a parquet file format.
1https://bitbucket.org/rr mstrs/nb ev paper 1/src/master/
Fig. 2: Data Analytical Workflow Implementation.
•Task (2): Then, the locations to parquet.py script is
executed to import raw station location data and integrate
multiple input files into one. The output is saved to a
parquet file format.
•Task (3): Next, the fuse location w events.py script is
triggered to fuse event data with the charging station
location information.
•Task (4): Focuses on recharge report event data in the
downstream analysis. The feat eng rech report.py script
is used to create new features (contextualized) based on
calculations involving existing data attributes and remove
events with a duration of 5 minutes or less (eliminating
11% of the raw records).
•Task (5): Extracts various partitions of the data using the
create batch ranges.py script to enable analysis accord-
ing to a particular week, month or season of the year.
•Task (6): Produces descriptive analytics artefacts such as
box plots and line charts using the descriptive.py script.
•Task (7): Transforms event data from a row-based charg-
ing event summary format to a row-based, per-minute
energy usage format using the ts rech report.py script.
•Task (8): Re-samples event data with different aggrega-
tion periods using the resample ts rech report.py script.
Half-hour (for weekly clustering jobs), one-hour (for
monthly clustering jobs) and four-hour (for seasonal
clustering jobs).
•Task (9): Create a column-based, per-station and period
energy usage format by pivoting the tables created in step
(8) by executing the format for agglo clustering kw.py
script.
•Task (10): Executes the agglo clustering kw.py script
to run the clustering algorithm using different input data
partitions (e.g. weekly, monthly, seasonal).
Using different feature spaces, we can effortlessly run
multiple experiments in our analytical workflow. Three exper-
iments have been selected to evaluate our proposed analytical
platform and their results are discussed in the next section (see
Section V-B).
V. RESULTS A ND DISCUSSION
A. Descriptive Analytical Results
A descriptive analysis was performed in order to get a global
understanding of the data. Fig. 3 summarizes the total monthly
energy transfer to vehicles for both types of charging stations.
The highest kWh month was August 2019 for both station
types (L2 stations : 3,147.21 kWh, L3 stations 16,923.89
kWh). From Fig. 4 we observe, as would be expected, the
aggregated monthly number of minutes spent charging at L2
stations is consistently higher than the same metric observed
for L3 stations. From Fig. 5, we can observe the month with
the highest number of L2 charging events was the month
of January 2020. The peak month for L3 charging event
frequency was August 2019, The high-frequency months for
both station types were July-August and December-February,
which largely corresponds to the summer and winter holiday
seasons respectively.
Fig. 3: Monthly kWh APR-2019 to APR-2020
Fig. 4: Monthly Duration Minutes APR-2019 to APR-2020
In Figures 4 and 5 we observe a sharp decline in total
monthly charge duration and event counts that becomes no-
ticeable in February and continues into March. In March, the
Fig. 5: Monthly Charge Events APR-2019 to APR-2020
provincial government made the first of many COVID-19-
associated announcements related to school and other closures.
The box plots in Figs. 6a and 6b summarize the monthly
kWh for charging events that occurred during the study time
frame. The interquartile range (IQR) is calculated as the
difference between quartiles 3 and 1. A common definition
for an outlier is a value that is more than 1.5 times the
interquartile range below the first quartile or above the third
quartile. According to this definition, there were many outliers
in the kWh values. This was especially true for L2 charge
events. The points above or below the whiskers are values
which are considered to be outliers in Figs. 6a and 6b.
(a) L2 Stations (b) L3 Stations
Fig. 6: Monthly L2, L3 kWh APR-2019 to APR-2020
For both station types, most months had multiple outliers ex-
cept for the months of June and September. L3 charging event
kWh values have relatively less outliers when compared to L2
events. The charging events described in Figs. 3 through 6b
were generated at public charging stations in New Brunswick,
which are mapped in Fig. 7.
Public charging stations were strategically located through-
out the province in order to allow for comfortable EV travel
distances on the road network with ample access to charging
opportunities. DC fast chargers (or L3 station outlets) gener-
ally transfer more energy to vehicles in a shorter amount of
time when compared to L2 charging. This is apparent when
comparing Tables II and III. The median monthly kWh values
are consistently and significantly lower for L2 charging events.
There is significant variability in L2 charge event kWh
values. The spread in these values is apparent when comparing
the monthly mean and median values and when observing
max and min values (See Table II). Additionally, the standard
deviation is above 4.5 most months, which indicates that
generally, on average, kWh values are above 4.5 units away
(a) L2 Stations (b) L3 Stations
Fig. 7: Charging stations in New Brunswick, Canada.
from the mean. A similar pattern of variability can be observed
in L3 charging events (See Table III).
TABLE II: L2 Energy Transfer Basic Statistics - (kWh)
Y-M N Mean Med Std Dev Min Max
2019-APR 471 4.12 1.31 6.48 0.15 42.71
2019-MAY 375 5.08 1.53 7.59 0.05 67.55
2019-JUN 257 6.94 4.54 7.09 0.09 42.56
2019-JUL 424 7.26 4.64 8.27 0.0 58.87
2019-AUG 433 7.27 5.22 7.87 0.09 60.26
2019-SEP 266 7.66 5.35 7.33 0.0 35.79
2019-OCT 409 5.59 1.9 7.45 0.12 54.26
2019-NOV 619 3.45 0.5 6.15 0.13 54.62
2019-DEC 622 3.57 1.5 6.63 0.07 70.55
2020-JAN 837 2.63 0.36 4.69 0.01 37.62
2020-FEB 707 2.36 0.32 4.36 0.01 32.92
2020-MAR 377 2.45 0.37 4.4 0.03 30.13
2020-APR 108 2.22 0.32 3.38 0.16 12.18
TABLE III: L3 Energy Transfer Basic Statistics - (kWh)
Y-M N Mean Med Std Dev Min Max
2019-APR 125 15.93 12.75 10.89 1.76 69.86
2019-MAY 185 18.48 15.08 12.26 1.44 72.87
2019-JUN 254 20.64 17.76 12.75 1.32 66.92
2019-JUL 631 20.3 16.98 12.38 1.44 93.9
2019-AUG 788 21.48 18.66 12.42 1.82 76.61
2019-SEP 336 20.73 17.16 13.47 2.06 69.51
2019-OCT 255 19.35 15.13 13.3 1.65 60.36
2019-NOV 208 17.53 14.17 11.36 2.0 58.48
2019-DEC 239 19.01 15.22 13.35 1.78 59.99
2020-JAN 180 20.64 15.92 15.32 0.76 79.82
2020-FEB 194 18.8 16.13 13.15 0.96 66.88
2020-MAR 179 18.33 14.19 11.92 1.89 54.33
2020-APR 26 17.61 17.9 9.51 1.04 41.13
B. Diagnostics Analytical Results
This section highlights interesting outcomes from the cluster
analysis of recharge events occurring at L2 and L3 charging
stations across the province of New Brunswick for the study
period. As discussed in Section IV-C, charging event data was
partitioned using different time granularity and aggregation
schemes. The results highlighted in this section used agglom-
erative clustering, PCA filtering at 70% variance with a 30
minute kWh aggregation rate. Additionally, longitude/latitude
information was included as features for each station. Optimal
dendrogram cut-offs were determined using the Cali´
nski-
Harabasz [10] method. Fig. 8 plots the weekly number of
clusters observed for weekly time slices. From the figure, we
observe the number of clusters varied significantly over the
weekly periods and L3 charging stations generally had more
clusters than L2 stations.
Fig. 8: Weekly Cluster Counts - L2 and L3 Events
1) Level 2 Recharge Events
First, we focus on L2 charging events and comment on
clustering results for recharge events that occurred during the
week starting on May 06, 2019 and compare these results to
events that occurred in the week starting on August 05, 2019.
During the week of May 06th, the number of L2 recharge
events was 68. Comparatively, the number of L2 recharge
events in the week of August 05th was 104.
The dendrogram in Fig. 9 is a tree structure which contains
all possible clusterings of a data set. The optimal cut-off
determined using the Cali´
nski-Harabasz [10] method for the
week of May 06th was 2. This cut-off clustered stations in 12
groupings.
Fig. 9: Dendrogram - L2 Events - Week of May 06, 2019
The map in Fig. 10 provides a spatial view of the cluster
members for the selected week. We see that the largest
cluster (cluster 3 in green) is comprised of stations which are
generally located in the center of the province. Additionally,
cluster number 1 and 4 member stations seem to be located
near the edges of the province only. Fig. 11 plots average kWh
values for the top 3 largest clusters. As can be observed in this
figure, cluster 4 stations, which generated the highest average
kWh peaks, are located close to the province of Quebec and
the State of Main in the United States. Cluster 1 stations are at
the edge of Nova Scotia and Prince Edward Island. The largest
cluster (cluster 3) represents consistently lower average kWh
energy transfer patterns. Clusters 1 and 4 had relatively higher
averages. Of note however, is the overall low aggregated kWh
values for top cluster members for this week. The aggregated
kWh of all L2 stations for the week was 334.62 kWh. The
percentage of this total attributed to the top 3 largest cluster
member stations was 12%. The smaller clusters consisted of
member stations with higher aggregated kWh values.
Fig. 10: Cluster Map - L2 Events - Week of May 06, 2019
Fig. 11: L2 Events - Key Clusters - Week of May 06, 2019
We now focus on agglomerative clustering results for the
104, L2, recharge events that occurred during the week of
August 05th. The number of events for this period was signif-
icantly higher when compared to the count of events during
the week starting May the 06th. There were 65% more events
in the week of August 05th. This increase in recharging events
is reflective of the peak summer period which started in June
for both station types and persisted throughout August (See
Fig. 5). The top 3 largest clusters mapped in Fig. 12, generally
included more member stations and largely covered more of
the province to split it in two halves. There was one sub-
cluster, cluster number 10, which was spatially included in
cluster number 4, that included two stations situated near each
other. The optimal cut-off in the dendrogram for this week was
also 2, which resulted in 17 station groupings. The dendrogram
for this clustering experiment is not included here for brevity.
Fig. 13 plots average kWh values for the top 3 largest station
clusters observed on the week of August 05 th. As can be
observed in this figure and in Fig. 12, cluster 10 member
stations, which generated the highest average kWh peaks,
contains two member stations located close to each other.
Cluster 4 and 5 stations members generated similar average
kWh peaks and roughly split the province in half.
Fig. 12: Cluster Map - L2 Events - Week of August 05, 2019
The aggregated kWh values of top cluster members for this
week were a little higher than values for the week of May
06th. Aggregated kWh of all L2 stations for this week was
879.53 kWh. The percentage of this total attributed to the top
3 largest cluster member stations was 17%. Again, the smaller
clusters consisted of member stations with higher aggregated
kWh values for the week.
Fig. 13: L2 Events - Key Clusters - Week of August 05, 2019
2) Level 3 Recharge Events
As mentioned in the data transformation section (See IV-C),
the analysis workflow can support running agglomerative clus-
tering of weekly, monthly or seasonal batches. In this section,
we focus on L3 stations and comment on station clusters
generated from all recharge events that occurred during the
month of May 2019. There were 185 recharge events that
occurred at L3 stations during this month. The total amount of
energy transferred to vehicles during the period was 3419.22
kWh. The agglomerative clustering experiment partitioned the
L3 stations into 20 groupings. The largest cluster included 7
stations. All other clusters were comprised of single stations.
The largest cluster (cluster number 9) was characterized as
grouping stations which produced relatively low average kWh
peaks. The aggregated value of kWh for all member stations in
this cluster was 2% of the total for the month. As can be seen
from Fig. 14, cluster number 9 member stations are broadly
located in the top right quadrant of the province.
Fig. 14: Cluster Map - L3 Events - May, 2019
Fig. 15 plots average kWh values for the top 3 largest
clusters. From this figure, we can observe that the cluster 9
member stations, on average, transferred very low amounts of
energy to vehicles during the month.
Fig. 15: L3 Events - May, 2019
Exploratory data analysis such as the one provided in
Section V-A is a good start in acquiring an initial under-
standing of trends and visible patterns in a data set. However,
the cluster analysis presented in this section can provide
additional insights. Identifying under-utilized charging stations
and similar stations groupings can be very useful to operators.
C. Discussion
The high capital costs of setting up public charging in-
frastructure and the use of public funds to support vehicle
electrification necessitates robust informed decision making.
The analysis in this work revealed that for some periods,
sections of the province can be spatially divided into broad
groupings of stations according to their energy utilization. The
results highlighted in Section V demonstrate that agglomer-
ative clustering is effective at grouping low kWh recharge
stations together by considering spatial and temporal attributes.
However, not all clustering experiments generated immediately
observable and interesting results. Additionally, a manual in-
spection of clustering results revealed that stations with normal
or relatively higher usage patterns were often not included in
the same clusters. Additional clustering experiments with other
algorithms and the filtering of outliers and fleet stations may
provide additional insights by producing more compact and
well-separated clusters.
VI. CONCLUSION AND FUTURE RESEARCH WO RK
A broad EV adoption scenario will require adequate public
charging infrastructure. An understanding of EV charging
patterns at public charging stations is crucial to foster adoption
while managing costs and optimizing placement of charging
infrastructure. The contributions in this work include an auto-
mated analytical workflow that enables the analysis of energy
utilization patterns of public charging infrastructure using real
charging data from station operators in New Brunswick. The
outcomes of this research is believed to provide useful insights
in planning and expanding infrastructure allocation. Future
work will explore if state of charge can be used to make charge
duration prediction for L3 chargers and whether this will be
useful as a service to station operators and users. Additionally,
future work will include exploring charging patterns from the
point of view of user behavior.
VII. ACKN OWLEDGMENTS
The authors would like to thank the New Brunswick Power
Corporation for providing access to the EV charging data
used in this research. This research was partially supported
by the NSERC/Cisco Industrial Research Chair, Grant IRCPJ
488403-1.
REFERENCES
[1] EVAG, “An electric vehicle roadmap for new brunswick a discussion
document for public and stakeholder engagement,” 2016.
[2] P. Ashkrof, G. Homem de Almeida Correia, and B. van Arem, “Analysis
of the effect of charging needs on battery electric vehicle drivers’ route
choice behaviour: A case study in the Netherlands,” Transportation
Research Part D: Transport and Environment, vol. 78, p. 102206, Jan.
2020.
[3] E. Abotalebi, D. M. Scott, and M. R. Ferguson, “Why is electric
vehicle uptake low in atlantic canada? a comparison to leading adoption
provinces,” Journal of Transport Geography, vol. 74, pp. 289–298, 2019.
[4] S. L. Monaca and L. Ryan, “The State of Play in Electric Vehicle Charg-
ing Services: Global Trends with Insight for Ireland,” no. November,
2018.
[5] S. Hardman, A. Jenn, G. Tal, J. Axsen, G. Beard, N. Daina, E. Figen-
baum, N. Jakobsson, P. Jochem, N. Kinnear, P. Pl¨
otz, J. Pontes, N. Refa,
F. Sprei, T. Turrentine, and B. Witkamp, “A review of consumer prefer-
ences of and interactions with electric vehicle charging infrastructure,”
Transportation Research Part D: Transport and Environment, vol. 62,
pp. 508–523, July 2018.
[6] H. Cao and M. Wachowicz, “An edge-fog-cloud architecture of stream-
ing analytics for internet of things applications,” Sensors, vol. 19, no. 16,
p. 3594, 2019.
[7] C. Boettiger, “An introduction to docker for reproducible research,” ACM
SIGOPS Operating Systems Review, vol. 49, no. 1, pp. 71–79, 2015.
[8] M. Zaharia, M. Chowdhury, M. J. Franklin, S. Shenker, and I. Stoica,
“Spark: Cluster computing with working sets,” in HotCloud, pp. 10–10,
2010.
[9] M. Zaharia, A. Chen, A. Davidson, A. Ghodsi, S. A. Hong, A. Kon-
winski, S. Murching, T. Nykodym, P. Ogilvie, M. Parkhe, et al.,
“Accelerating the machine learning lifecycle with mlflow.,” IEEE Data
Eng. Bull., vol. 41, no. 4, pp. 39–45, 2018.
[10] T. Cali ´
nski and J. Harabasz, “A dendrite method for cluster analysis,”
Communications in Statistics-theory and Methods, vol. 3, no. 1, pp. 1–
27, 1974.
