This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Prediction of EV Charging Behavior using

Machine Learning
Sakib Shahriar1, A. R. Al-Ali1, (Life Senior Member, IEEE), Ahmed H. Osman2, (Senior
Member, IEEE), Salam Dhou1, (Member, IEEE), and Mais Nijim3, (Member, IEEE)
1
Computer Science and engineering, Department American University of Sharjah, Sharjah, UAE
2
Electrical Engineering Department, Department American University of Sharjah, Sharjah, UAE
3
Electrical Engineering and Computer Science Department, Texas A&M University – Kingsville, Kingsville, TX 78363 USA
Corresponding author: A. R. Al-Ali (e-mail: [email protected]).

ABSTRACT As a key pillar of smart transportation in smart city applications, electric vehicles (EVs) are
becoming increasingly popular for their contribution in reducing greenhouse gas emissions. One of the key
challenges, however, is the strain on power grid infrastructure that comes with large-scale EV deployment.
The solution to this lies in utilization of smart scheduling algorithms to manage the growing public charging
demand. Using data-driven tools and machine learning algorithms to learn the EV charging behavior can
improve scheduling algorithms. Researchers have focused on using historical charging data for predictions
of behavior such as departure time and energy needs. However, variables such as weather, traffic, and nearby
events, which have been neglected to a large extent, can perhaps add meaningful representations, and provide
better predictions. Therefore, in this paper we propose the usage of historical charging data in conjunction
with weather, traffic, and events data to predict EV session duration and energy consumption using popular
machine learning algorithms including random forest, SVM, XGBoost and deep neural networks. The best
predictive performance is achieved by an ensemble learning model, with SMAPE scores of 9.9% and 11.6%
for session duration and energy consumptions, respectively, which improves upon the existing works in the
literature. In both predictions, we demonstrate a significant improvement compared to previous work on the
same dataset and we highlight the importance of traffic and weather information for charging behavior
predictions.

INDEX TERMS Electric vehicles (EVs), charging behavior, machine learning, smart city, smart
transportation

I. INTRODUCTION reliability has grown and satisfaction among EV owners are

Climate change has become a growing concern in recent years higher [6]. The driver flexibility has also increased with the
with thirty-three countries jointly declaring a climate addition of charging stations in many parts of the world, often
emergency as of January 2021 [1]. Global energy lead by various government initiatives encouraging further
consumption is a major contributor to the climate crisis, and in adoption of EVs. These factors have placed EVs to be in a pole
particular, the transportation sector accounts for over a quarter position with regards to providing a clean source of
of the global energy consumption [2]. The United Nations transportation.
(UN) projects that two thirds of the world’s population will There still remains a few challenges, most notably the
reside in urban areas by 2050 [3]. This would increase the charging time and public charging needs, despite the
demand for urban mobility, leading to further energy promising potential. Although EV charging time has
consumption and emissions of greenhouse gases. Studies have significantly decreased over the years, it is still on average
shown that electric vehicles (EVs) have the potential to reduce much higher than the refueling time for ICE vehicles.
carbon emissions by 45% compared to conventional internal Emerging charging technologies such as extreme fast
combustion engine (ICE) vehicles [4]. EVs were initially charging [7] and wireless charging [8] are promising but are
limited by factors such as reliability and battery range, which still overcoming various challenges and will require years
have significantly improved in recent years and led to an before being adopted. The constraints from charging
increase in EV popularity [5]. As a result, the trust in EV

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

infrastructure means that most EV owners rely on public regression models. Historical charging data was utilized, and
charging stations, which poses a strain on power distribution eight features were used including, car ID, car type,
grid due to the high-power requirements of the EVs [9]. To weekday, charging point, car park location, parking floor and
avoid power grid degradation and failures, un-coordinated arrival time. For prediction, three regression models were
charging behavior must be avoided. The optimal solution is trained namely, linear regression, XGBoost and artificial
to better manage the scheduling of charging stations. The neural network (ANN). XGBoost achieved the best results
research on smart scheduling using data driven approaches with mean absolute error (MAE) of 82 minutes. In [24],
are plentiful and include optimization [10] and metaheuristic ensemble machine learning using SVM, random forest (RF)
[11] approaches. Furthermore, psychological factors and diffusion-based kernel density estimator (DKDE) was
influencing charging behavior [12] as well as transactions used for session length and energy consumption predictions.
data and interviews with EV drivers [13] have been used for For training, historical charging records from two separate
charging behavior analysis. A comprehensive review of datasets were used, with one of them being public and the
charging behavior analysis using machine learning and data- other being residential charging. The ensemble model
driven approaches is presented in [14], which concludes that performed better than the individual models in both
machine learning based approaches are more suitable to predictions and the reported SMAPEs were 10.4% for
scheduling approaches with the ability to provide duration and 7.5% for the consumption.
quantification and more realistic representation.
TABLE I
SUMMARY OF RELATED WORKS
A. RELATED WORKS
Source Prediction Model Features Results
Although predictions of EV charging behavior can have [21] Session GMM Historical SMAPE:
various categories, the focus of this work will be on session length, charging 14.4%
duration and energy consumption. Examples of other energy data duration,
consumption 15.9%
charging behavior include the prediction of whether the EVs consumption
will be charged the next day [15], identification of the use of [22] Arrival time, SVM Historical MAPE:
fast charging [16], prediction of the time to next plug [17], departure charging 2.9%
time data arrival,
charge profile prediction [18], charging speed prediction 3.7%
[19] and prediction of charging capacity and the daily departure
charging times [20]. These behaviors provide valuable [23] Departure XGBoost Historical MAE: 82
insights, but the prediction of session duration and energy time charging minutes
data, vehicle
time is more valuable for scheduling purposes. type,
As will be defined in the following sections, session charging
duration is directly related to the departure time. It is the location
[24] Session Ensemble Historical SMAPE:
arrival time, which is a known variable, minus the departure length, model of charging 10.4%
time. Therefore, one can assume the prediction of either the energy SVM, RF data duration,
session duration or the departure time to have the same consumption & DKDE 7.5%
consumption
application. Lee et al. [21] introduced a novel dataset for [25] Start time, Linear Historical -
non-residential EV charging consisting of over 30000 session regression charging
charging sessions. They used gaussian mixture models length, data
(GMM) to predict session duration and energy needs by energy
consumption
considering the distribution of the known arrival times. The [26] Energy XGBoost Historical R2: 0.52,
testing dataset included the month of December 2018 and the requirements charging MAE: 4.6
reported symmetric mean absolute percentage errors data, season, kWh
weekday,
(SMAPEs) were 14.4% and 15.9% for the session duration location
and energy consumption, respectively. In this work, only type,
historical charging data was considered for obtaining the charging
fees
predictions. In [22], the authors used support vector [27] Energy k-NN Last few SMAPE:
machines (SVM) for the prediction of arrival and departure consumption days energy 15.3%
time for EV commuters in a university campus. Using consumption
historical arrival and departure times and temporal features [28] Energy PSF Last few SMAPE:
consumption days energy 14.1%
i.e., week, day, and hour, the reported mean absolute consumption
percentage error (MAPE) was 2.9% and 3.7% for arrival and
departure times, respectively. For comparison, a simple Xiong et al. [25] predicted the start time and session
persistence model was used as reference and SVM duration using mean estimation. Session duration was then
hyperparameter tuning was not addressed in the work. used to obtain energy consumption predictions using linear
Frendo et al. [23] predicted the departure time of EVs using regression. The charging behavior predictions were integrated

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

to flatten the charging load profile and stabilize the power grid. discusses the results of this work. Future research directions
However, the prediction performances were not evaluated are provided in Section V, and Section VI concludes the paper.
quantitatively. In [26], several regression models were used to
II. BACKGROUND
predict the energy requirements from public charging stations This section summarizes the background information
data for the US state of Nebraska. Besides historical charging including the algorithms used in this work and the evaluation
data, parameters such as season, weekday, location type and metrics for predictions.
charging fees were used as input features. On the test set,
XGBoost model outperformed linear regression, RF and SVM A. SUPERVISED MACHINE LEARNING
obtaining a R2 score of 0.52 and MAE of 4.6 kWh. The authors The main objective in machine learning (ML) is to develop a
in [27] used k-nearest neighbor (k-NN) to predict the energy learning framework that can learn from experience, i.e., the
consumption at a charging outlet using data from a university training dataset, without explicit programming. Primarily, ML
campus. The problem was formulated as time-series forecast algorithms are classified as either supervised learning or
whereby energy consumption prediction for the next day (next unsupervised learning. In unsupervised ML, the training data
24 hours) was made using energy consumption of previous is not labeled, and the goal of the algorithm is to group similar
days. The highest SMAPE was 15.3% using k value of 1 (1- data points. Conversely, in supervised learning, the models are
NN) and a time-weighted dot product dissimilarity measure. trained from labeled dataset that contains the specified output
Similarly, Majidpour et al. [28] also predicted the next day or target variable, i.e., the variable to be predicted. The
energy needs of a charging station based on previous days representation between the input and target variable is learned
energy consumption using various algorithms including SVM iteratively by optimizing a specific objective function. In this
and RF. They also experimented with pattern sequence-based work, the target variables, i.e., the session duration and the
forecasting (PSF) [29], where clustering is first applied to energy consumption are both labeled, and thus supervised
classify the days and predictions are made for that day. The learning will be used. Furthermore, since both target variables
PSF-based approach provided the most accurate results with are continuous values, we are going to use regression models
average SMAPE value of 14.1%. Table 1 provides a summary as opposed to classification models which deals with
of the related works in the literature. categorical target values. The four regression models used in
this work are RF, SVM, XGBoost and deep ANN. The
B. OBJECTIVES following paragraphs describe each of them briefly.
Although the above works from the literature have A decision tree (DT) can be used to separate complex
successfully applied machine learning for the prediction of decisions into a combination of simpler decisions using split
session duration and energy consumption, they have mainly points from the input features. Leaf nodes are the points
focused on utilizing historical charging data. In some cases, where no further split is made whereas a decision node is the
additional derived features such as vehicle information, point where decisions take place. Predictions are made by
charging location information and seasonal information were taking the average value of all the items in the leaf node in
used. This has motivated us in this work to investigate the regression. Although simple to implement, a single DT is
use of additional input features including weather, traffic and prone to overfitting. To overcome this problem, multiple
local events and observe its impact on the accuracy of
DTs can be aggregated, and this is the essence of a random
charging behavior predictions. The key contributions of this forest (RF) algorithm. Bagging method is used in this case
work are the following: where the trees are created from various bootstrap sample
1) We propose a novel approach in EV charging behavior which is sample with replacement. The average value of the
prediction that utilizes weather, traffic, and local predictions across all the trees are taken as the final
events data along with historical charging records. prediction for regression problems [30].
2) We use several machine learning algorithms including Similar to a RF, a gradient boosting algorithm [31] makes
RF, SVM, XGBoost and ANN for predictions of
use of multiple DTs. However, in this algorithm each tree is
session duration and energy consumption on the built sequentially and as a result the errors made by previous
adaptive charging network (ACN) dataset.
trees are taken into consideration which often leads to
3) We empirically show that the use of additional data has
superior performance. XGBoost [32] is a more recent
a positive impact on the accuracy of predictions and
variation of the gradient boosting algorithm. XGBoost has
significantly improves upon the previous work on the
gained popularity over the last few years for its success in
same dataset that used only historical charging
machine learning competitions mainly due to it being
information.
effective in dealing with the bias-variance tradeoff [33]. This
The rest of the paper is organized as follows. Background
means that the algorithm is able to avoid overfitting on the
information including key concepts in machine learning is
training data while at the same time maintaining enough
provided in Section II. This is followed by a detailed
complexity to obtain meaningful representations.
explanation of the methodology, including dataset description,
and experimental setup in Section III. Section IV presents and

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

A support vector machine (SVM) [34] is used for both commonly used in related works. Equations (1)-(4) defines the
classification and regression problems. It is sometimes metrics that will be used in this work:
referred to as support vector regression [35] when
Root mean square error (RMSE):
exclusively applied to regression problems. SVM separates
the classes with the best hyperplane that can maximize the "
margin between the respective classes. The key objective is ∑#!$%)𝑦! − 𝑦̅! -
𝑅𝑀𝑆𝐸 = ' (1)
to map the inputs to high dimensional feature spaces where 𝑛
they are linearly separable. This is achieved using kernels
such as linear, polynomial, and radial basis function (RBF). Mean absolute error (MAE):
SVM is not suitable for larger datasets due to its long training #
time. 1
𝑀𝐴𝐸 = 12𝑦! − 𝑦̅! 2 (2)
Deep learning-based models contain a large amount of 𝑛
!$%
composition of learned functions. Using layered hierarchy of
concepts, complex concepts are defined in terms of simpler Coefficient of determination or R2:
concepts and more abstract representations are gathered "
using less abstract ones [36]. Variations of deep learning "
∑#!$%)𝑦! − 𝑦̅! -
𝑅 = 1− # (3)
algorithms include convolutional and recurrent neural ∑!$%(𝑦! − 𝜇)"
networks, which have been successful in image and audio
classification tasks. In this work, we consider artificial neural Symmetric mean absolute percentage error (SMAPE):
networks (ANN), often referred to as a multilayer perceptron #
(MLP). MLPs utilize non-linear approximation given a set of 1 2𝑦! − 𝑦̅! 2
𝑆𝑀𝐴𝑃𝐸 = 1 ∗ 100% (4)
input features and can be used for both regression and 𝑛 )|𝑦! | + 2𝑦̅! 2-/2
!$%
classification. An MLP consists of input layer which is fed
with a given set of input features, the hidden layers which
learns the representations and the output layer which makes where 𝑦 represents the actual value, 𝑦̅ is the predicted value,
the final predictions. When the number of hidden layers is 𝜇 is the average of the actual values and 𝑛 represents the
two or more, the model is referred to as deep ANN. groups of values in the dataset. Generally, lower scores of
In ensemble learning, set of individually trained classifiers RMSE, MAE and SMAPE indicate accurate predictions, and
are combined and then used to predict new instances, often this occurs when the predicted value, 𝑦̅ is very close to the
providing more accurate predictive performance than the actual value 𝑦. The R2 value is a measure of goodness of fit for
individual classifiers [37]. Figure 1 illustrates the concept of regression and is usually a score between 0 and 1. A score of
ensemble learning. Both RF and XGBoost are examples of 1 indicates perfect predictions and generally a higher value
ensemble learning, where individual models (in these cases represents better performance. We do not consider mean
DTs) are first evaluated and then integrated into a single absolute percentage error because it is inconvenient when the
model. The motivation behind such approach is similar to actual value 𝑦 is close to 0, therefore creating a bias. Rather
asking multiple experts about an opinion, and then taking their we consider SMAPE which is more suitable for EV charging
votes to make the final decision [38]. prediction because both the original and the predicted values
are in the denominator [24].

III. METHODOLOGY
In this section, we define the approach used for the prediction
of charging behavior. We formulate the problem, describe the
dataset, highlight the preprocessing steps, and discuss the
methods for training the learning models.

A. EV CHARGING BEHAVIOR
Assuming 𝑡&'# represents the connection time when the car
first plugs in, 𝑡(!)&'# represents the disconnection time when
Figure 1. Illustration of ensemble learning the car plugs out and leaves the station and 𝑒 represents the
energy delivered to the car during the session, we consider the
B. EVALUATION OF REGRESSION MODELS session charging behavior 𝐵)*))!'# as following:
To assess the performance of predictions made by regression
models, numerous metrics are used as discussed in [39]. In this 𝐵)*))!'# ≜ (𝑡&'# , 𝑡(!)&'# , 𝑒) (5)
work, we will define and use four measures that were

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

Based on the above, we can define the length of charging Pasadena (where the charging data originates from) provides
session or the session duration, 𝑆(+, , as follows: an open data site [45] for the traffic count around the city.
However, for most roads in the city it contains traffic count for
𝑆(+, = 𝑡(!)&'# − 𝑡&'# (6)
some period of time and therefore is not usable in our case
where we require regular interval data. Additionally, not all
roads and streets are covered. As a result, we decided to use
In this work, we predict both the session duration and the
traffic data from google maps, which has also been used in
session energy consumption of an individual charging record
previous machine learning applications [46]. The data is
and assume that the connection time is known.
collected by recording the location data from the commuter’s
mobile devices provided they use the application and have
B. DATASET DESCRIPTION
agreed to share their location. The data collected from
Besides the charging dataset, we also make use of weather,
individuals is anonymized and aggregated to address any
traffic, and local events data in order to predict the charging
privacy concerns [47]. The google maps distance matrix API
behavior. We will briefly describe the datasets used and
can be used to retrieve the data. Given a source and destination
highlight their attributes.
coordinates, the travel distance and the time taken is returned
Scheduling of EV charging is more significant in public
for a given departure time. We retrieved historical trip time for
charging structures due to the unpredictable nature of the
9 of the closest roads and streets which one must take to access
charging behavior, especially in places like shopping malls.
the charging station.
The ACN [21] dataset is among the few publicly available
Since the charging station is located in the Caltech
datasets for non-residential EV charging and will be utilized
university campus, we decided to include campus events and
in this work. The dataset contains charging records from two
find out if the number of events have an impact on the charging
stations in the university campus, namely JPL and Caltech.
behavior. The number of events in an hour were obtained from
Unlike the Caltech station, which is open to public, the JPL
the Caltech university website calendar [48]. For
station is only accessible to employees and therefore will not
simplification, we decided to round the minutes to the nearest
be considered in this work. Registered users can manually
hour, therefore if an event started at 10.20 am, it was counted
enter additional details, such as their estimated departure
as an event starting at 10 am.
time and requested energy, by scanning a QR code through
their mobile applications. The dataset can be accessed from C. DATA PREPROCESSING
[40] by either a web portal or python application Cleaning and preprocessing the dataset is vital to ensuring
programming interface (API). the quality of the predictive models. These include removing
Although there is a small weather station located at the faulty records and outliers.
Caltech campus [41], we did not consider it for this work due The presence of outliers can negatively impact the model
to missing values and irregular interval recordings for the wind performance. A common technique of graphically detecting
variable. Additionally, this station did not record variables outliers is boxplots [49]. The boxplots for both target
such as rainfall and snowfall which could potentially impact variables contained outliers, as shown in Figure 2. We notice
charging behavior. We therefore used the weather data from that the outliers for both variables are not consistent, i.e., we
NASA’s Modern-Era Retrospective analysis for Research and have far too many outlier points for energy consumption than
Applications, Version 2 (MERRA-2) [42] which provides data the session duration. It is possible that certain vehicles
for the precise location of the charging station. The accuracy consume far greater amount of energy even if the session
of satellite weather data in comparison to ground stations has duration is not too long.
been compared in [43]. Although it has been shown that given
a specific location some weather parameters may be more
accurately detected using ground stations, for the purpose of
this work we do not require a high level of accuracy but rather
a more general perception of the impact of weather on
charging behaviors. For example, we are interested in
observing how the charging behavior is impacted during
heavy rainfall as opposed to drier conditions.
Obtaining historical traffic data for specific roads and
regions is challenging. Conventional traffic collection Figure 2. Boxplots of energy consumption (left), session duration (right)
methods include intrusive approaches such as road tubes and
piezoelectric sensors and non-intrusive approaches including As a result, we opted to perform multivariate outlier
microwave radar and video image detection [44]. With most detection using the isolation forest algorithm which constructs
of these approaches, scalability is an issue, and in most cases, an ensemble of iTrees for a given data set. The outliers are
specific roads are not covered. For instance, the city of those instances which have short average path lengths on the

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

iTrees [50]. By randomly selecting a variable and a split value information. Instead of simply selecting the traffic level for a
between the minimum and maximum of the selected variable, given time, we selected the total traffic after arrival until the
the observations are ‘isolated’. Partitioning of observations are end of the day. If a vehicle arrived at 2 pm, for instance, we
repeated recursively until all of them have been isolated. After accumulated the traffic from 2 pm until the end of that day.
the partitioning, observations that have shorter path lengths for This would allow the model to learn how the traffic level
some particular points are likely to be the outliers. Figure 3 impact the charging behavior. Similarly, we considered the
illustrates the process in detecting the outlier of the target total events after arrival until the end of the day.
variables. A total of 697 outliers were detected which accounts
for 4% of the total observations. D. FEATURE ENGINEERING
Feature engineering refers to the transformation of data into
meaningful representation using human knowledge. This
process is labor intensive but important nonetheless as this is
a weakness of the learning algorithms. Feature engineering
relies on human ingenuity and prior knowledge to
compensate for the inability of the algorithms to extract and
organize the discriminative information from the data [53].
We discuss the future engineering steps next.
Firstly, we convert the time fields that will be used by the
models into numeric format by simply dividing the minute by
60 and adding to the hour. Then, for each charging record, we
find out their average departure time, session duration and
energy consumption. This is done by finding out the user ID
of the charging record and aggregating his previous records.
We use the arrival time as a numeric feature. However, the
arrival time also has other components such as the date
information. Using this, we extract the hour of the day, day of
the month, month of the year, day of the week, whether the
day is a weekend and whether the day falls in a US federal
Figure 3. Outlier detection using isolation forest holiday. However, temporal information such as day, hour,
and month are cyclic ordinal features. This is because the hour
For the charging data, we only considered charging records value of 23 corresponding to 11 pm, for example, is actually
that were registered, i.e., contained user IDs, and this close to the hour value of 0 which corresponds to 12 am. To
accounted for 97% of the records. For the weather data, the represent the proximity of these values, trigonometric
time of recording was in universal time and we used the pytz transformation is performed as following:
[51] library in python to convert the time zone to be the same
as that of the charging records. We also converted the 𝑓- = sin)2𝜋𝑓/𝑚𝑎𝑥(𝑓)- (7)
temperature units from kelvin to degrees Celsius. Then for
each given hour, we also computed average of the previous 7
hours of weather and the average of the next 10 hours. This 𝑓. = 𝑐𝑜𝑠)2𝜋𝑓/𝑚𝑎𝑥(𝑓)- (8)
would allow us to understand how the previous weather and
the weather after charging impacts charging decision. For
instance, heavy snowfall in the previous hours may account where 𝑓 represents the cyclic feature to be transformed, 𝑓- and
for shorter charging duration and so on. We also had to convert 𝑓. represents the first and second components of the cyclic
the time zone from coordinated universal time for the traffic feature, respectively. To transform other categorical variables,
data. We then aggregated the traffic for each hour across the one-hot encoding was used, where a single variable with n
nine selected roads and streets. It must be noted that we points and k distinct classes is transformed into k binary
considered the average trip time as well as the maximum trip variables with n points each. For numeric variables, feature
time as estimated by google maps. Finally, we aggregated the scaling is a common transformation where the goal is to
total events in the campus for each hour. normalize the range of the numeric features. There are various
To merge the various data, the time-series fields were scaling techniques, including scaling by domain where all the
converted to date-time objects using pandas [52] library. Then features are scaled to a specific range such as [0, 1] and scaling
to obtain weather, traffic, and events for a particular charging to minmax where the features are scaled to the range [0, R], in
record, we first obtained the nearest hour that the connection which case the minimum of the maximum value of feature in
time belongs to. For example, the connection time of 22:11 all directions is assigned as the radius of the sphere R [54].
belongs to 10 pm. This allows us to easily extract the other However, in this work we have used standardization which

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

ensures the values of each feature to have zero mean and unit seasonal factors into consideration during training. The dataset
variance. The transformations were performed using the was split such that 80% of the records were used for model
preprocessing package of the Scikit-learn [55] library. Table training and 20% for evaluation. During the training phase, we
2 lists the features used for training. performed K-fold cross validation where the algorithms are
repeatedly trained K times with a fraction 1/K training
TABLE II examples left out for testing [56]. In this case, we selected the
LIST OF FEATURES AND THEIR DESCRIPTIONS
Feature Description
common K value of 10. To determine model hyperparameters,
session_length Length of charging duration, target we utilized grid search method which determines the optimal
variable set of parameters from a given list by trying out all possible
kWh_delivered Session energy consumption, target values of the specified parameters [57]. We performed the grid
variable search across K-folds, selected to be 5 in this case to speed up
time_con Numerical representation of the
connection time (arrival time)
the grid search. We then evaluated all the models using the
day_of_week Day of the week, one-hot encoded aforementioned regression metrics. Inspired by the success of
is_weekend Binary variable indicating whether the ensemble learning methods in previous works, we also
session took place in a weekend decided to experiment with ensemble learning. We used two
holiday Binary variable indicating whether the variants of ensemble stacking, namely voting regressor and
session took place on a US federal holiday
hr_x, hr_y Sine and Cosine components of the hour
stacking regressor, using the ensemble package of the Scikit-
day_x, day_y Sine and Cosine components of the day learn library. In a voting regressor, several base regressors are
mnth_x, mnth_y Sine component of the month trained on the entire training set, and the average predictions
mean_d_time Historical average departure time made by the base models are treated as the final prediction.
mean_con Historical average consumption Stacking regressor is based on the concept of stacked
mean_dur Historical average session length
traffic_aft_arvl average traffic level after arrival
generalization where predictions made by the base models are
max_traffic_aft_arvl maximum traffic level after arrival used as inputs to a final estimator, which is trained using cross-
events_after_arrival total campus events after arrival validation, to generate predictions [58]. Figure 4 provides a
avg_temp_prv average temperature of last 7 hours graphical representation of the framework.
avg_temp_nxt average temperature of next 10 hours
avg_hum_prv average humidity of last 7 hours IV. RESULTS AND DISCUSSION
avg_hum_nxt average humidity of next 10 hours
We begin the experiment with RF algorithm which can be
avg_win_prv average wind speed of last 7 hours
avg_win_nxt average wind speed of next 10 hours used to visualize the variable importance [30]. This is a
avg_rain_prv average rainfall of last 7 hours method for feature selection where certain variables that are
avg_rain_nxt average rainfall of next 10 hours not important and can often hinder performance are removed.
avg_snwfall_prv average snowfall of last 7 hours In this case, the inclusion of the least important variables had
avg_snwfall_nxt average snowfall of next 10 hours a very insignificant performance increase and hence we
avg_snwdpth_prv average snow depth of last 7 hours
avg_snwdpth_nxt average snow depth of next 10 hours
decided to include them in model training. Additionally,
avg_irradiation_prv average irradiation of last 7 hours variables can be ranked in terms of their relative importance.
avg_irradiation_nxt average irradiation of next 10 hours This is determined by each feature’s contribution in
determining the most effective splits. In Figures 5 and 6, we
plot the top 10 important variables for session duration and
E. MODEL SELECTION AND EXPERIMENTAL SETUP energy consumption, respectively. The two most important
We selected all charging sessions from the ACN dataset that predictors of session duration are the maximum traffic after
belonged to the 2019 calendar year, which ensures we take the arrival and the time of connection. This indicates the

Figure 4. Graphical representation of the proposed framework

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

usefulness of including traffic information for the prediction Therefore, we aggregated the 3 best performing models in
of session duration. However, for energy consumption, the the training phase into 2 ensemble models, which resulted in
historical average consumption is by far the most significant. improved cross validation scores. Next, we present the
This is because a specific vehicle will consume similar energy results on the test set. For reference, we also selected the user
if the session duration is consistent. estimates of their departures as prediction. This value was
collected through a smart phone app where users were asked
to enter their estimates of their departure time and
consumption upon arrival. We summarize the results on the
test set in Table 4.

TABLE IV
TEST SCORES FOR SESSION DURATION
Metrics/Model RMSE MAE R2 SMAPE
(mins) (mins) (%)
RF 98.7 68.0 0.63 10.1
Figure 5. Top ten features for session duration SVM 101 67.4 0.64 10.1
XGBoost 97.9 68.0 0.63 10.1
Deep ANN 101 73.7 0.57 10.9
Voting Ensemble 97.7 66.5 0.73 9.92
Stacking Ensemble 97.5 67.1 0.73 9.95
User predictions 430 394 -4.20 69.9

As highlighted, the best results are obtained using the

ensemble learning approach, which is consistent with
previous works [15], [24]. Voting regressor performs best on
2 metrics and stacking regressor performs the best in terms
Figure 6. Top ten features for energy consumption of RMSE, whereas they both achieve the same R2 score. The
results are consistent with the training performance with RF,
A. SESSION DURATION PREDICTIONS SVM and XGBoost resulting in similar performance and
As mentioned earlier, the hyperparameters for the models deep ANN performs the worst of the four base models.
were determined using the grid search approach. For the deep Predictions made by user about their own session length is
ANN training, we experimentally determined an architecture also far off the actual session length. This indicates that
with 3 hidden layers of 64, 32 and 16 nodes respectively to perhaps relying on users to provide an estimate of their own
be the most suitable. Rectified linear units (Relu) [59] was departure time is perhaps not suitable.
used as the activation function for all hidden layers and the
output layer contained a linear activation as we are expecting B. ENERGY CONSUMPTION PREDICTIONS
the prediction to be a numeric value. The learning rate value Similar approach to the session duration prediction was also
was set to 0.001 and we used the Adam [60] algorithm for used here. The only exception was the deep ANN
model optimization. The training batch size was 32 and the architecture which in this case contained 2 hidden layers with
number of iterations were 15 epochs. Appendix 1 displays 64 and 16 nodes, respectively. The training batch size was
the training loss curve and Table 3 summarizes the 10-fold 64 and the number of epochs was set to 20. Appendix 2
cross validation scores on the training set. presents the loss curve from the training phase. Table 5
summarizes the 10-fold cross validation scores on the
TABLE III training set.
TRAINING SCORES FOR SESSION DURATION TABLE V
Metrics/Model RMSE MAE R2 SMAPE TRAINING SCORES FOR ENERGY CONSUMPTION
(mins) (mins) (%) Metrics/Model RMSE MAE R2 SMAPE
(kWh) (kWh) (%)
RF 100 68.7 0.72 10.1 RF 5.49 3.40 0.69 11.9
SVM 102 68.2 0.71 10.2 SVM 5.65 3.53 0.67 12.6
XGBoost 5.56 3.49 0.68 12.4
XGBoost 100 69.3 0.72 10.3
Deep ANN 5.61 3.60 0.67 12.9
Deep ANN 103 73.1 0.71 10.7 Voting Ensemble 5.50 3.42 0.69 12.0
Voting Ensemble 99.5 67.4 0.72 10.0 Stacking Ensemble 5.48 3.40 0.69 11.9
Stacking Ensemble 99.5 68.0 0.72 10.1
RF has the best cross validation scores whereas the other
3 models have similar scores. We selected the top 3 models,
The training scores are very similar for RF, SVM and
i.e., RF, SVM and XGBoost to form the 2 ensemble models.
XGBoost whereas deep ANN performs slightly worse.
In this case, the ensemble models did not improve upon the

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

best performing RF model but rather achieved similar results [28]). We summarize the results from the previous works in
on training. The results from the test set are presented in comparison to the one achieved in this work in Table 6. In
Table 6. We also compare the results with user predictions comparison to [24], the results obtained in this work for
about their consumptions. session duration is more accurate although we do not
The best results as highlighted were obtained using the improve upon their results for energy consumption. This is
stacking ensemble model. The improvement using ensemble most likely because the authors in [24] utilized both
learning for energy consumption prediction was perhaps not residential and non-residential data for their predictions, and
as significant when we compare with the session duration. residential charging behavior in most cases are more
The user predictions about their consumptions are not consistent. However, it must be noted that all previous works
accurate in this case as well. except [21] used a different dataset to this work and therefore
a comparison is perhaps not suitable. Therefore, keeping the
TABLE VI comparison across the same dataset, we can conclude that the
TEST SCORES FOR ENERGY CONSUMPTION
Metrics/Model RMSE MAE R2 SMAPE
utilization of the additional weather, traffic and events data
(kWh) (kWh) (%) resulted in an improvement in the EV charging behavior
RF 5.50 3.39 0.54 11.7 predictions.
SVM 5.69 3.54 0.51 12.4
XGBoost 5.61 3.48 0.51 12.1 TABLE VI
Deep ANN 5.65 3.55 0.55 12.5 PERFORMANCE COMPARISON WITH PREVIOUS WORKS
Voting Ensemble 5.54 3.41 0.69 11.8 Source Session Energy Dataset Used
Stacking Ensemble 5.50 3.38 0.70 11.6 Duration Consumption
User predictions 20.6 11.8 0.04 55.0 [21] SMAPE: SMAPE: 15.9% ACN (historical
14.4% charging)
C. COMPARISON AND DISCUSSION [23] MAE: 82 Not considered German charging
minutes data (historical
When we compare across both predictions, looking at the
charging, vehicle &
overall R2 and the SMAPE, it appears that the prediction of location info)
energy consumption is perhaps more difficult. This is [24] SMAPE: SMAPE: 7.5% UCLA campus
consistent with the previous work on the ACN data [21]. 10.4% (historical charging)
However, in another case the opposite was observed [24], and Residential
charging data from
i.e., the prediction of energy consumption was easier. UK
Moreover, in both scenarios, it was also noticed that the user [26] Not R2: 0.52 Nebraska public
predictions about their own behavior is very different to their considered charging (historical
actual behavior, which further emphasizes the need for charging, temporal
& location)
predictive analytics. The users’ predictions in terms of their [27] Not SMAPE: 15.3% UCLA campus
energy consumption are slightly more accurate when considered (historical energy)
compared to their predictions of session duration as indicated [28] Not SMAPE: 14.1% UCLA campus
by better R2 and SMAPE values. This could be due to the considered (historical energy)
Our SMAPE: SMAPE: ACN, weather,
users’ lack of interest in entering their estimates every time
work 9.92%, 11.6%, R2: 0.7 traffic, and events
they decide to charge their vehicles. We also noticed that the MAE: 66.5 data
performance using deep ANN was the least accurate in both minutes
cases. Although deep learning models are proven superior in
dealing with images and audio data where feature extraction
is not performed, in applications such as this where we V. RECOMMENDATIONS AND FUTURE WORK
perform feature extraction, traditional ML models usually We have quantitatively shown in the previous section that the
perform better. Furthermore, predictions made by ensemble traffic and weather data were important predictors in EV
learning outperformed predictions made by individual ML charging behavior, particularly in the case of session duration.
models in both scenarios, although the impact was more Although the use of local events data (campus events in this
significant for session duration prediction. This is most likely case) had insignificant impact in terms of performance gain, it
because in the first scenario, the top 3 performing models had cannot be ruled out for future work. In this work, we obtained
similar training performance and combining their predictions all campus events from the university calendar. However,
resulted in an improvement. However, in the latter scenario, perhaps only the significant events that draw more crowd
one model clearly outperformed the rest in training and should be taken into consideration. It is possible that events
hence the improvement using ensemble learning was not data may not impact predictions in a university campus
significant. setting. However, for other public spaces such as shopping
Looking at the previous works in the literature, the results malls for example, events like end of the year sale could be
in this work outperformed all the previous works that important predictors. Therefore, similar experiments on other
reported similar evaluation metrics ([21], [23], [26], [27], public charging spaces should be carried out to find the impact

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

of local events. Social media can also be explored as a means Program from the American University of Sharjah, UAE. This
to obtain information about local events as well as driver paper represents the opinions of the authors and does not mean
behavior. For instance, social media has been shown to be a to represent the position or opinions of the American
good tool for estimating human behavior [61] and also is a University of Sharjah.
significant predictor of truck drivers’ travel time [62]. It is also
likely that the use of vehicle information such as the vehicle REFERENCES
model and vehicle type can improve predictions, especially in [1] “Climate Emergency Declaration and Mobilisation In Action,”
CEDAMIA. https://www.cedamia.org/global/ (accessed Jan. 17, 2021).
terms of energy consumption. Some of the previous works [2] “Key World Energy Statistics 2018 – Analysis,” IEA.
have utilized vehicle information [23] but not in conjunction https://www.iea.org/reports/key-world-energy-statistics-2019 (accessed
with weather, traffic, and events. Finally, to better understand Jun. 01, 2020).
[3] “68% of the world population projected to live in urban areas by
the charging behavior during the COVID-19 pandemic, a case 2050, says UN,” UN DESA | United Nations Department of Economic and
study should be conducted using the proposed approach to Social Affairs, May 16, 2018.
validate the predictive performance in uncertain situations. https://www.un.org/development/desa/en/news/population/2018-revision-
of-world-urbanization-prospects.html (accessed Jun. 01, 2020).
[4] X. Zhang, F. Gao, X. Gong, Z. Wang, and Y. Liu, “Comparison
VI. CONCLUSION of Climate Change Impact Between Power System of Electric Vehicles and
In this work, we presented a framework for the prediction of Internal Combustion Engine Vehicles,” in Advances in Energy and
two of the most important EV charging behaviors with regards Environmental Materials, Singapore, 2018, pp. 739–747.
[5] “Global EV Outlook 2019 – Analysis,” IEA.
to scheduling, namely EV session duration and energy https://www.iea.org/reports/global-ev-outlook-2019 (accessed Jun. 01,
consumption. Unlike previous work, we utilized weather, 2020).
traffic, and events data along with the historical charging data. [6] Y. Kwon, S. Son, and K. Jang, “User satisfaction with battery
electric vehicles in South Korea,” Transp. Res. Part Transp. Environ., vol.
We trained four popular ML models along with two ensemble 82, p. 102306, 2020, doi: https://doi.org/10.1016/j.trd.2020.102306.
learning algorithms for the prediction of charging behavior. [7] D. Ronanki, A. Kelkar, and S. S. Williamson, “Extreme Fast
The results obtained in terms of prediction performance is Charging Technology—Prospects to Enhance Sustainable Electric
superior to the results in the previous works. We have also Transportation,” Energies, vol. 12, no. 19, 2019, doi: 10.3390/en12193721.
[8] S. A. Q. Mohammed and J.-W. Jung, “A Comprehensive State-
provided a significant improvement of charging behavior of-the-Art Review of Wired/Wireless Charging Technologies for Battery
prediction on the ACN dataset and demonstrated the potential Electric Vehicles: Classification/Common Topologies/Future Research
of utilizing traffic and weather information in charging Issues,” IEEE Access, vol. 9, pp. 19572–19585, 2021, doi:
10.1109/ACCESS.2021.3055027.
behavior prediction. [9] A. Ramanujam, P. Sankaranarayanan, A. Vasan, R. Jayaprakash,
V. Sarangan, and A. Sivasubramaniam, “Qantifying The Impact of Electric
APPENDIX Vehicles On The Electric Grid: A Simulation Based Case-Study,” in
Proceedings of the Eighth International Conference on Future Energy
Systems, New York, NY, USA, 2017, pp. 228–233. doi:
10.1145/3077839.3077854.
[10] M. Ş. Kuran, A. Carneiro Viana, L. Iannone, D. Kofman, G.
Mermoud, and J. P. Vasseur, “A Smart Parking Lot Management System
for Scheduling the Recharging of Electric Vehicles,” IEEE Trans. Smart
Grid, vol. 6, no. 6, pp. 2942–2953, 2015, doi: 10.1109/TSG.2015.2403287.
[11] J. García-Álvarez, M. A. González, and C. R. Vela,
“Metaheuristics for solving a real-world electric vehicle charging
scheduling problem,” Appl. Soft Comput., vol. 65, pp. 292–306, 2018, doi:
https://doi.org/10.1016/j.asoc.2018.01.010.
[12] L. Hu, J. Dong, and Z. Lin, “Modeling charging behavior of
battery electric vehicle drivers: A cumulative prospect theory based
approach,” Transp. Res. Part C Emerg. Technol., vol. 102, pp. 474–489,
Appendix 1. Validation loss curve for session duration
2019, doi: https://doi.org/10.1016/j.trc.2019.03.027.
[13] J. Spoelstra, “Charging behaviour of Dutch EV drivers,”
Master’s Thesis, 2014.
[14] S. Shahriar, A. R. Al-Ali, A. H. Osman, S. Dhou, and M. Nijim,
“Machine Learning Approaches for EV Charging Behavior: A Review,”
IEEE Access, vol. 8, pp. 168980–168993, 2020, doi:
10.1109/ACCESS.2020.3023388.
[15] S. Ai, A. Chakravorty, and C. Rong, “Household EV Charging
Demand Prediction Using Machine and Ensemble Learning,” in 2018 IEEE
International Conference on Energy Internet (ICEI), 2018, pp. 163–168.
[16] Y. Yang, Z. Tan, and Y. Ren, “Research on Factors That
Influence the Fast Charging Behavior of Private Battery Electric Vehicles,”
Sustainability, vol. 12, no. 8, 2020, doi: 10.3390/su12083439.
Appendix 2. Validation loss curve for energy consumption [17] S. Venticinque and S. Nacchia, “Learning and Prediction of E-
Car Charging Requirements for Flexible Loads Shifting,” in Internet and
ACKNOWLEDGMENT Distributed Computing Systems, Cham, 2019, pp. 284–293.
[18] O. Frendo, J. Graf, N. Gaertner, and H. Stuckenschmidt, “Data-
The work in this paper was supported, in part, by the Computer driven smart charging for heterogeneous electric vehicle fleets,” Energy AI,
Science and Engineering department and the Open Access vol. 1, p. 100007, 2020, doi: https://doi.org/10.1016/j.egyai.2020.100007.

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

[19] J. J. Mies, J. R. Helmus, and R. Van den Hoed, “Estimating the [42] R. Gelaro et al., “The Modern-Era Retrospective Analysis for
Charging Profile of Individual Charge Sessions of Electric Vehicles in The Research and Applications, Version 2 (MERRA-2),” J. Clim., vol. 30, no.
Netherlands,” World Electr. Veh. J., vol. 9, no. 2, 2018, doi: 14, pp. 5419–5454, Jul. 2017, doi: 10.1175/JCLI-D-16-0758.1.
10.3390/wevj9020017. [43] R. Mendelsohn, P. Kurukulasuriya, A. Basist, F. Kogan, and C.
[20] Y. Lu et al., “The Application of Improved Random Forest Williams, “Climate analysis with satellite versus weather station data,”
Algorithm on the Prediction of Electric Vehicle Charging Load,” Energies, Clim. Change, vol. 81, no. 1, pp. 71–83, Mar. 2007, doi: 10.1007/s10584-
vol. 11, no. 11, 2018, doi: 10.3390/en11113207. 006-9139-x.
[21] Z. J. Lee, T. Li, and S. H. Low, “ACN-Data: Analysis and [44] G. Leduc and others, “Road traffic data: Collection methods and
Applications of an Open EV Charging Dataset,” in Proceedings of the Tenth applications,” Work. Pap. Energy Transp. Clim. Change, vol. 1, no. 55, pp.
ACM International Conference on Future Energy Systems, New York, NY, 1–55, 2008.
USA, 2019, pp. 139–149. doi: 10.1145/3307772.3328313. [45] “Pasadena Traffic Count Website.”
[22] Z. Xu, “Forecasting Electric Vehicle Arrival & Departure Time https://data.cityofpasadena.net/datasets/eaaffc1269994f0e8966e2024647cc
On UCSD Campus using Support Vector Machines,” PhD Thesis, UC San 56 (accessed Jan. 21, 2021).
Diego, 2017. [46] F. Goudarzi, “Travel Time Prediction: Comparison of Machine
[23] O. Frendo, N. Gaertner, and H. Stuckenschmidt, “Improving Learning Algorithms in a Case Study,” in 2018 IEEE 20th International
Smart Charging Prioritization by Predicting Electric Vehicle Departure Conference on High Performance Computing and Communications; IEEE
Time,” IEEE Trans. Intell. Transp. Syst., pp. 1–8, 2020. 16th International Conference on Smart City; IEEE 4th International
[24] Y.-W. Chung, B. Khaki, T. Li, C. Chu, and R. Gadh, “Ensemble Conference on Data Science and Systems (HPCC/SmartCity/DSS), 2018,
machine learning-based algorithm for electric vehicle user behavior pp. 1404–1407. doi: 10.1109/HPCC/SmartCity/DSS.2018.00232.
prediction,” Appl. Energy, vol. 254, p. 113732, 2019, doi: [47] D. Barth, “The bright side of sitting in traffic: Crowdsourcing
https://doi.org/10.1016/j.apenergy.2019.113732. road congestion data,” Google Off. Blog, 2009.
[25] Y. Xiong, C. Chu, R. Gadh, and B. Wang, “Distributed optimal [48] “Calendar | www.caltech.edu.”
vehicle grid integration strategy with user behavior prediction,” in 2017 https://www.caltech.edu/campus-life-events/master-calendar (accessed Jan.
IEEE Power Energy Society General Meeting, 2017, pp. 1–5. 23, 2021).
[26] A. Almaghrebi, F. Aljuheshi, M. Rafaie, K. James, and M. [49] J. W. Tukey, Exploratory data analysis, vol. 2. Reading, Mass.,
Alahmad, “Data-Driven Charging Demand Prediction at Public Charging 1977.
Stations Using Supervised Machine Learning Regression Methods,” [50] F. T. Liu, K. M. Ting, and Z. Zhou, “Isolation Forest,” in 2008
Energies, vol. 13, no. 16, p. 4231, 2020. Eighth IEEE International Conference on Data Mining, 2008, pp. 413–422.
[27] M. Majidpour, C. Qiu, P. Chu, R. Gadh, and H. R. Pota, “Fast doi: 10.1109/ICDM.2008.17.
Prediction for Sparse Time Series: Demand Forecast of EV Charging [51] S. Bishop, Pytz–world timezone definitions for Python.
Stations for Cell Phone Applications,” IEEE Trans. Ind. Inform., vol. 11, Online][Cited: May 25, 2012.] http://pytz. sourceforge. net, 2016.
no. 1, pp. 242–250, 2015. [52] W. McKinney and others, “pandas: a foundational Python library
[28] M. Majidpour, C. Qiu, P. Chu, R. Gadh, and H. R. Pota, “A novel for data analysis and statistics,” Python High Perform. Sci. Comput., vol.
forecasting algorithm for electric vehicle charging stations,” in 2014 14, no. 9, pp. 1–9, 2011.
International Conference on Connected Vehicles and Expo (ICCVE), 2014, [53] Y. Bengio, A. Courville, and P. Vincent, “Representation
pp. 1035–1040. Learning: A Review and New Perspectives,” IEEE Trans. Pattern Anal.
[29] N. Bokde, M. W. Beck, F. [Martínez Álvarez, and K. Kulat, “A Mach. Intell., vol. 35, no. 8, pp. 1798–1828, 2013, doi:
novel imputation methodology for time series based on pattern sequence 10.1109/TPAMI.2013.50.
forecasting,” Pattern Recognit. Lett., vol. 116, pp. 88–96, 2018, doi: [54] P. Juszczak, D. Tax, and R. P. Duin, “Feature scaling in support
https://doi.org/10.1016/j.patrec.2018.09.020. vector data description,” in Proc. asci, 2002, pp. 95–102.
[30] A. Liaw, M. Wiener, and others, “Classification and regression [55] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,”
by randomForest,” R News, vol. 2, no. 3, pp. 18–22, 2002. J. Mach. Learn. Res., vol. 12, pp. 2825–2830, 2011.
[31] J. H. Friedman, “Greedy Function Approximation: A Gradient [56] Y. Bengio and Y. Grandvalet, “No unbiased estimator of the
Boosting Machine,” Ann. Stat., vol. 29, no. 5, pp. 1189–1232, 2001. variance of k-fold cross-validation,” J. Mach. Learn. Res., vol. 5, no. Sep,
[32] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting pp. 1089–1105, 2004.
System,” in Proceedings of the 22nd ACM SIGKDD International [57] Qiujun Huang, Jingli Mao, and Yong Liu, “An improved grid
Conference on Knowledge Discovery and Data Mining, New York, NY, search algorithm of SVR parameters optimization,” in 2012 IEEE 14th
USA, 2016, pp. 785–794. doi: 10.1145/2939672.2939785. International Conference on Communication Technology, 2012, pp. 1022–
[33] D. Nielsen, “Tree boosting with xgboost-why does xgboost win" 1026. doi: 10.1109/ICCT.2012.6511415.
every" machine learning competition?,” Master’s Thesis, NTNU, 2016. [58] D. H. Wolpert, “Stacked generalization,” Neural Netw., vol. 5,
[34] C. J. C. Burges, “A Tutorial on Support Vector Machines for no. 2, pp. 241–259, 1992, doi: https://doi.org/10.1016/S0893-
Pattern Recognition,” Data Min. Knowl. Discov., vol. 2, no. 2, pp. 121–167, 6080(05)80023-1.
Jun. 1998, doi: 10.1023/A:1009715923555. [59] V. Nair and G. E. Hinton, “Rectified linear units improve
[35] M. Awad and R. Khanna, “Support Vector Regression,” in restricted boltzmann machines,” 2010.
Efficient Learning Machines: Theories, Concepts, and Applications for [60] D. P. Kingma and J. Ba, “Adam: A method for stochastic
Engineers and System Designers, M. Awad and R. Khanna, Eds. Berkeley, optimization,” ArXiv Prepr. ArXiv14126980, 2014.
CA: Apress, 2015, pp. 67–80. doi: 10.1007/978-1-4302-5990-9_4. [61] M.-A. Abbasi, S.-K. Chai, H. Liu, and K. Sagoo, “Real-world
[36] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. The behavior analysis through a social media lens,” in International Conference
MIT Press, 2016. on Social Computing, Behavioral-Cultural Modeling, and Prediction, 2012,
[37] D. Opitz and R. Maclin, “Popular ensemble methods: An pp. 18–26.
empirical study,” J. Artif. Intell. Res., vol. 11, pp. 169–198, 1999. [62] D. Yuniar, L. Djakfar, A. Wicaksono, and A. Efendi, “Truck
[38] R. Polikar, “Ensemble based systems in decision making,” IEEE driver behavior and travel time effectiveness using smart GPS,” Civ. Eng.
Circuits Syst. Mag., vol. 6, no. 3, pp. 21–45, 2006. J., vol. 6, no. 4, pp. 724–732, 2020.
[39] A. Botchkarev, “A new typology design of performance metrics
to measure errors in machine learning regression algorithms,” Interdiscip.
J. Inf. Knowl. Manag., vol. 14, pp. 45–76, Jan. 2019.
[40] “ACN-Data -- A Public EV Charging Dataset.”
https://ev.caltech.edu/dataset (accessed Jul. 02, 2020).
[41] “TCCON Weather.” http://tccon-weather.caltech.edu/ (accessed
Jul. 02, 2020).

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3103119, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

University of Sharjah, Sharjah, UAE. His research interests and activities

Sakib Shahriar received his B.S. (2018) in include power system analysis and power system protection.
Computer Engineering from American University
of Sharjah, U.A.E. He is currently pursuing his Salam Dhou (Member, IEEE) received the
M.S. degree in Computer Engineering at the B.Sc. and M.Sc. degrees in computer science
American University of Sharjah, U.A.E. He is also from the Jordan University of Science and
a research assistant at the same university. His Technology, Irbid, Jordan, in 2004 and 2007,
research interests include Machine Learning, Big respectively, and the Ph.D. degree in electrical
Data Analytics and Deep Learning. and computer engineering from Virginia
Commonwealth University, Richmond, VA,
USA in 2013. She worked as a Postdoctoral
Research Fellow in the Department of Radiation
A. R. Al-Ali received the Ph.D. in electrical Oncology at Harvard Medical School, Boston,
engineering and a minor in computer science Massachusetts, USA, between 2013 and 2016.
from Vanderbilt University, Nashville, TN, USA, She is currently an Assistant Professor in the Department of Computer
1990; MS from Polytechnic Institute of New Science and Engineering, and the Biomedical Engineering Graduate
York, USA, 1986 and B. Sc.EE from Aleppo Program with the American University of Sharjah, Sharjah, United Arab
University, Syria, 1979. From 1991-2000, he was Emirates. Her research interests include machine learning and data mining,
a faculty with the EE Dept. at KFUPM, Saudi computer vision, and medical imaging and informatics.
Arabia. Since 2000–present, he is working as
professor of computer engineering with American
Mais Nijim received her B.S in Computer Science
University of Sharjah, UAE. His research,
teaching interests include embedded systems, cyber physical systems, IoT from Princess Sumaya University in Amman,
and IIoT applications in smart cities. Jordan. She received her M.S. degree in Computer
Science from New Mexico State University in 2004,
Ahmed H. Osman received the B.Sc. and M.Sc. and her Ph.D. in Computer Science from New
degrees in electrical engineering from Helwan Mexico Institute of Mining and Technology in 2007.
University, Cairo, Egypt, in 1991 and 1996, She has been an Associate Professor with the Department of Electrical
respectively, and the Ph.D. degree in electrical
engineering from the University of Calgary, Engineering and Computer Science, Texas A&M University-Kingsville. Her
Calgary, AB, Canada, in 2003. From 2004 to research interests include Machine Learning, Cyber Physical System,
2008, he was an Assistant Professor with the Cybersecurity, and Wireless Sensor Network. She is the editor of the
Department of Electrical and Computer
International Journal of Sensor Network.
Engineering, University of Calgary. He is
currently a Professor at the Department of Electrical Engineering, American

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/