Accurate Recommendation of EV Charging Stations Driven by

Availability Status Prediction

Meriem Manai

1 a

, Bassem Sellami

2 b

and Sadok Ben Yahia

3 c

LIPAH-LR11ES14, Faculty of Sciences of Tunis, University of Tunis, El Manar, Tunis, Tunisia

Tallinn University of Technology, Tallinn, Estonia

University of Southern Denmark, Sønderborg, Denmark

Keywords:

Electric Vehicles, Charging Stations, Machine Learning, Prediction, Real-Time Availability.

Abstract:

The electric vehicle (EV) market is experiencing substantial growth, and it is anticipated to play a major role as

a replacement for fossil fuel-powered vehicles in transportation automation systems. Nevertheless, as a rule of

thumb, EVs depend on electric charges, where appropriate usage, charging, and energy management are vital

requirements. Examining the work that was done before gave us a reason and a basis for making a system

that forecasts the real-time availability of electric vehicle charging stations that uses a scalable prediction

engine built into a server-side software application that can be used by many people. The implementation

process involved scraping data from various sources, creating datasets, and applying feature engineering to

the data model. We then applied fundamental models of machine learning to the pre-processed dataset, and

subsequently, we proceeded to construct and train an artiﬁcial neural network model as the prediction engine.

Notably, the results of our research demonstrate that, in terms of precision, recall, and F1-scores, our approach

surpasses existing solutions in the literature. These ﬁndings underscore the signiﬁcance of our approach in

enhancing the efﬁciency and usability of EVs, thereby signiﬁcantly contributing to the acceleration of their

adoption in the transportation sector.

1 INTRODUCTION

Amidst growing concerns regarding environmental is-

sues and the detrimental impacts of fossil fuels, the

global adoption of electric vehicles (EVs) is witness-

ing a signiﬁcant surge (El Halim, A. et al., 2022). De-

spite challenges posed by the COVID-19 pandemic,

the EV industry has demonstrated remarkable re-

silience, with global sales exceeding 10 million units

in 2022 (Intekin, 2022). China leads this trend, hold-

ing over 50% of the global EV market share and

showcasing a strong commitment to sustainable mo-

bility (International Energy Agency, 2023). However,

the widespread adoption of EVs heavily relies on a

robust and accessible charging infrastructure (Tam-

bunan et al., 2023). Research indicates that the avail-

ability and quality of charging stations signiﬁcantly

inﬂuence EV adoption rates, with inadequate infras-

tructure posing a barrier even in countries with high

https://orcid.org/0009-0005-5919-6233

https://orcid.org/0000-0001-6869-3518

https://orcid.org/0000-0001-8939-8948

home charging prevalence (El-Fedany et al., 2023;

Karike et al., 2023; Schulz and Rode, 2022; Gian-

soldati et al., 2020; Engel et al., 2018). There is a link

between the number of public charging points and the

number of EVs that are bought, especially in cities

(Hennlock and WP4 Shift, 2020). This shows how

important it is to have a widespread and standardized

charging infrastructure to help EVs grow and become

more popular (George et al., 2022; Balakrishnan and

Pillai, 2023).

Applying machine learning and deep learning

techniques has become increasingly prevalent in opti-

mizing various aspects of electric vehicle (EV) charg-

ing infrastructure. Soldan et al. (Soldan, 2021) em-

ployed data stream analysis and logistic regression

to predict short-term EV charging station occupancy,

utilizing both batch and real-time data. Sao et al.(Sao

et al., 2021) introduced the Deep Fusion of Dynamic

and Static Information Models (DFDS), integrating

static and dynamic data patterns to enhance the accu-

racy of charging station occupancy forecasting. Hecht

et al. (Hecht et al., 2021) explored ensemble machine

learning methods, speciﬁcally Gradient Boosting De-

Manai, M., Sellami, B. and Ben Yahia, S.

Accurate Recommendation of EV Charging Stations Driven by Availability Status Prediction.

DOI: 10.5220/0012752600003753

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Conference on Software Technologies (ICSOFT 2024), pages 351-358

ISBN: 978-989-758-706-1; ISSN: 2184-2833

351

cision Trees (GBDTs) and Random Forest Classiﬁers,

to predict EV charging station accessibility. Mean-

while, Ma and Faye (Ma and Faye, 2022) utilized

LSTM neural networks for intraday predictions of

public charging station occupancy, incorporating di-

verse data sources to improve accuracy. These studies

demonstrate the potential of advanced algorithms in

optimizing EV charging infrastructure and predicting

station availability, while also highlighting the need

for further research to address real-world complexi-

ties and diverse charging scenarios.

In comparing these studies, Soldan et al.’s (Sol-

dan, 2021) data streaming approach offers real-time

updates, which could be advantageous for dynamic

charging environments. Still, their model’s perfor-

mance might differ with varying data quality and size.

Sao et al.’s (Sao et al., 2021) DFDS model demon-

strated the potential of combining static and dynamic

information, but its application in diverse geographi-

cal regions remains to be tested. Hecht et al.’s (Hecht

et al., 2021) ensemble approach showcased promis-

ing accuracy, but the effectiveness might vary with

different datasets and charging station types. Finally,

the LSTM model proposed by Ma et al. (Ma and

Faye, 2022) showed promise in intraday forecasts,

but its generalizability to more extensive and diverse

datasets warrants further investigation. These studies

provide valuable insights into EV charging station oc-

cupancy forecasting, but additional research is neces-

sary to validate and reﬁne their methodologies under

various real-world scenarios.

Beyond predicting EV charging station availabil-

ity, research has also focused on understanding charg-

ing behaviors and forecasting future charging demand

using machine learning and deep learning methodolo-

gies. Qiao and Lin (Qiao and Lin, 2021) developed

predictive models to anticipate upcoming charging

demand, capturing the behavior patterns of both long-

term and short-term users. Their study employed

XGBoost, SVR, and GBDT algorithms on real-world

charging data, with XGBoost demonstrating superior

performance. Shahriar et al. (Shahriar et al., 2021)

investigated EV charging behavior, session duration,

and energy consumption, addressing concerns about

power grid strain from high-power charging. Their re-

search utilized various machine learning algorithms,

including DANN, Random Forest, and XGBoost, on

the ACN dataset (Lee et al., 2019). These studies pro-

vide valuable insights into EV charging behavior and

demand patterns, contributing to developing efﬁcient

charging strategies and infrastructure planning.

Research on EV charging behavior and demand

forecasting by Qiao and Lin (Qiao and Lin, 2021) and

Shahriar et al. (Shahriar et al., 2021) reveals valu-

able insights into user patterns and potential grid im-

pacts. While XGBoost proves to be a robust predic-

tion tool in both studies, further research is needed to

develop comprehensive models that consider diverse

user behaviors, charging patterns, and regional varia-

tions, ensuring applicability across different contexts

and addressing potential power grid challenges.

The studies conducted by Almaghrebi et al. (Al-

maghrebi et al., 2020), Kim et al. (Kim and Kim,

2021), Zhao et al. (Zhao et al., 2021), and Zhu et al.

(Zhu et al., 2019b; Zhu et al., 2019a) contributed to

the research on EV charging demand forecasting.

Almaghrebi et al. constructed a model to fore-

cast the charging demand using various machine-

learning methods and evaluated its performance on

a real-world dataset (Almaghrebi et al., 2020). In

their study, Kim et al. compared different modeling

approaches to predict charging demand, using both

historical data and outside factors (Kim and Kim,

2021). Zhao et al. proposed a data-driven frame-

work addressing overﬁtting issues with limited data

in complex environments. The model was evalu-

ated using real-world EV data and compared to high-

performance models (Zhao et al., 2021). Zhu et al.

developed an EV charging demand prediction algo-

rithm using deep learning techniques and found that

LSTM outperformed traditional time-series forecast-

ers (Zhu et al., 2019b; Zhu et al., 2019a).

The research by Zhao et al. (Zhao et al., 2021),

Almaghrebi et al. (Almaghrebi et al., 2020), Kim et

al. (Kim and Kim, 2021), and Zhu et al. (Zhu et al.,

2019b; Zhu et al., 2019a) helps us predict how much

EV charging will be needed, but there are some prob-

lems with the models they use, how easy they are to

understand, how stable they are, and how they can

be used for long-term planning. Further research is

necessary to reﬁne these approaches and ensure their

effectiveness across diverse scenarios.

In conclusion, the studies collectively contribute

to advancing EV charging demand prediction. But

making models easier to understand, testing how well

they work in different environments, and thinking

about both short-term and long-term prediction hori-

zons would make these approaches more useful and

effective(Intekin, 2022).

Despite the availability of various machine learn-

ing and deep learning models for analyzing EV charg-

ing infrastructure, a research gap exists in accurately

predicting real-time charging station availability. This

gap, coupled with challenges related to model accu-

racy and scalability, hinders effective charging plan-

ning for EV drivers. As the number of EVs and

charging stations increases, accurate forecasting be-

comes increasingly crucial to ensure drivers can con-

ICSOFT 2024 - 19th International Conference on Software Technologies

352

veniently locate available stations that meet their im-

mediate charging needs.

Forecasting electric vehicle charging stations is

crucial for planning power systems and transportation

networks. Existing research has made progress in sus-

tainable transportation infrastructure using statistical

models (Gruosso et al., 2020; Kim and Kim, 2021)

and machine learning algorithms (Hecht et al., 2022;

Yi et al., 2022; Koohfar et al., 2023) to predict de-

mand and optimal station placement. However, a re-

search gap still needs to be ﬁlled in forecasting station

availability, facing challenges of accuracy and scala-

bility due to limited data and complex user behavior

and charging patterns.

As a result, this paper tackles the challenge of

accurately predicting EV charging station availabil-

ity by developing a system that utilizes web scraping,

machine learning, and deep learning models. The re-

sulting prediction system, integrated into a distributed

software system, provides real-time availability infor-

mation to EV drivers, addressing the limitations of

existing research and enhancing the overall charging

experience. These activities are carried out among

those of the Horizon Europe ENERGETIC project

https://energeticproject.eu/the-project/, where

we design a three-layer resilient framework for BMS

advanced analytics. The prediction service for EV

charging station availability is among the ones de-

ployed in the fog layer.

The rest of the paper is structured as follows: Sec-

tion 2 presents our proposed methodology. The out-

comes of our experimental evaluation are discussed

in Section 3. Finally, Section 4 summarizes our take-

away messages and alludes to future issues.

2 METHODOLOGY

In this section, we introduce our methodology for

predicting the status of electric vehicle (EV) charg-

ing stations using an artiﬁcial neural network (ANN)

model. The model uses Python code with the Keras

deep learning library, employing a sequential archi-

tecture. The rationale behind using an ANN lies in its

ability to effectively capture complex non-linear rela-

tionships between input features and charging station

status categories. The ANN’s versatility as a universal

function approximator allows it to handle diverse and

intricate patterns in the charging station data, mak-

ing it well-suited for the multiclass classiﬁcation task.

The dataset used for training and evaluation is sourced

from a CSV ﬁle containing comprehensive informa-

tion about various charging stations that was provided

in (Intekin, 2022). The model’s performance is as-

sessed using several evaluation metrics, including F1-

score, precision, and recall. Additionally, visualiza-

tion functions are employed to depict the prediction

results in an interpretable manner.

This research aims to develop a robust system for

predicting EV charging station availability by utiliz-

ing web scraping, machine learning, and deep learn-

ing techniques. By analyzing historical data and iden-

tifying patterns, the system accurately forecasts the

availability of charging spots, providing EV drivers

with real-time information through a user-friendly in-

terface. This approach empowers drivers and en-

hances their overall charging experience.

In the following, we will provide a detailed ac-

count of each step mentioned above, including data

scraping, dataset generation, ML/DL models for pre-

diction, and EV charging spot availability prediction.

Figure 1 illustrates the key stages of the pro-

posed EV charging station forecasting model. To

extract meaningful information for precise predic-

tions, the process starts with data scraping from var-

ious sources. Machine learning algorithms, includ-

ing KNN, logistic regression, random forest, and sup-

port vector machines, are then employed to train ro-

bust predictive models. This iterative training process

enables the model to identify patterns and relation-

ships within the data, ultimately leading to reliable

real-time predictions of charging station availability.

This model highlights the intricate interplay between

data scraping, feature engineering, and model train-

ing and underscores their collective contribution to

the ultimate goal of providing reliable predictions for

EV charging station accessibility. To put the issue in

Figure 1: The overall architecture of the proposed approach.

simpler terms, Figure 2 illustrates how the model is

designed to forecast the availability rate for charging

stations along the driver’s route and present this infor-

mation to the driver. Notably, the third station stands

Accurate Recommendation of EV Charging Stations Driven by Availability Status Prediction

353

out with the highest availability rate, making it the top

recommendation for the driver.

Figure 2: Model simpliﬁcation.

Data on EV charging stations was collected

through web scraping from two sources: OpenData

Paris ”Belib” (bel, a), providing information on 1, 822

stations in Paris, France, and Eneﬁt Volt, covering

over 185 stations across Estonia. Belib’s public API

(bel, b) facilitated real-time data retrieval, while Ene-

ﬁt Volt required a different approach due to the lack of

a public API. The collected data, stored in CSV ﬁles,

included station status, location, timestamps, and ad-

ditional attributes. This comprehensive dataset served

as the foundation for model development and training

(scr, ).

The second data source examined in this study

pertains to a privately owned Estonian company

known as Eneﬁt Volt. The company asserts that it

operates the largest network of EV charging stations

in Estonia, with over 185 stations spread throughout

the country. The company offers an array of chargers

tailored to various requirements and is continuously

expanding its network (ene, ). Nonetheless, as Eneﬁt

Volt does not provide a public API, scraping this data

source involved a somewhat distinct approach com-

pared to the previous source. As with the script used

for the ﬁrst data source, the response from Eneﬁt Volt

contains details of multiple charging sites, which are

then processed one by one to retrieve comprehensive

data for each particular charging station. The col-

lected data for each charging station is then stored in

a CSV ﬁle, resulting in over 370, 000 rows of data,

each representing a single EV charging point with the

charging points information (Intekin, 2022).

The data collection timestamp is also recorded us-

ing datetime.now() for each entry. Once the data is

scraped from both sources, the rows are added to a

CSV ﬁle to create the ﬁnal dataset for constructing

the model.

To apply the generated datasets in machine learn-

https://opendata.paris.fr/explore/dataset/belib-point

s-de-recharge-pour-vehicules-electriques-disponibilite

-temps/api/?disjunctive.statut\ pdc\&disjunctive.arrondis

sement\&disjunctive.arrondissement

ing algorithms, it is crucial to perform both feature

engineering and pre-processing on the datasets. Fea-

ture engineering encompasses transforming raw data

into useful features for the model, including select-

ing the most pertinent predictor variables. The model

comprises both outcome and predictor variables (Fea,

). Due to two distinct datasets with various data types

and columns, two separate preprocessors were devel-

oped for each of them. Our second step is preprocess-

ing the obtained data to prepare it for ANN model

training. The charging station information is sourced

from the CSV ﬁle,” and relevant data manipulations

are performed using the Pandas library. The charging

station status labels are encoded into numerical val-

ues, ensuring compatibility with the ANN’s numer-

ical computations. Subsequently, the encoded out-

put variable is one-hot encoded, transforming it into a

categorical form suitable for multiclass classiﬁcation

tasks. Feature extraction is accomplished by separat-

ing the “Status” column from the dataset. Lastly, the

dataset is divided into training and testing sets using

the train test split function, ensuring proper separa-

tion for model evaluation.

The core of our methodology lies in designing the

model architecture to effectively capture the intricate

relationships between the input features and the out-

put charging station status. The sequential model ar-

chitecture offered by Keras is adopted to design the

ANN. The model comprises two essential layers: a

dense hidden layer with 256 neurons and a dense out-

put layer with eight neurons. The hidden layer em-

ploys the Rectiﬁed Linear Activation (ReLU) func-

tion, enabling the model to extract meaningful fea-

tures from the input data. The ReLU activation in-

troduces non-linearity, allowing the ANN to handle

complex patterns that may exist in the charging sta-

tion data. The output layer utilizes the softmax acti-

vation function, providing class probabilities for each

charging station status category. The softmax activa-

tion is particularly advantageous for multiclass clas-

siﬁcation, as it allows the model to express its conﬁ-

dence in predicting different status labels.

Once the model architecture is established, the

next step is to train the model on the prepared train-

ing dataset. The training is performed over 20 epochs

with a batch size of 10 to optimize model parameters

and ensure convergence. The choice of 20 epochs

for training provides ample iterations for the model

to learn from the data and adjust its internal weights.

The Stochastic Gradient Descent (SGD) optimizer

achieves efﬁcient optimization during training with a

learning rate of 0.3. We ﬁnally reached the predic-

tion needed to obtain predicted values for the test-

ing dataset, enabling further analysis and comparison

ICSOFT 2024 - 19th International Conference on Software Technologies

354

with actual charging station statuses.

In conclusion, the methodology revolves around

using an Artiﬁcial Neural Network model to predict

the status of electric vehicle charging stations. The

ANN’s capability to capture non-linear patterns and

its adaptability as a universal function approximator

make it suitable for the multiclass classiﬁcation task.

3 RESULTS AND DISCUSSION

This section presents the results of evaluating the per-

formance of multi-class classiﬁcation machine learn-

ing and deep learning models. The performance of

these models is assessed using precision, recall, and

F1 score metrics to measure and compare their ef-

fectiveness in predicting charging station availability.

Table 1 displays the outcomes obtained through the

machine learning/deep learning (ML/DL) methodol-

ogy applied to the Paris dataset. The results demon-

strate that among the ML baseline models, Random

Forest (RF) achieved the highest performance across

all metrics, with an accuracy rate exceeding 95%. K-

Nearest Neighbors (KNN) followed closely with a

rate of 94% in all metrics. Regarding the F1 Score,

the Support Vector Machine (SVM) yielded the low-

est rate at 59.54%, with Logistic Regression closely

behind at 60.73%.

However, the artiﬁcial neural network (ANN) model,

despite slightly trailing behind the K-nearest neigh-

bors (KNN) and random forest (RF) models in per-

formance scores, managed to attain a precision of

89.46%, a recall of 87.94%, and an F1 score of

88.69%.

Based on the results presented in Table 1 and Fig-

ure 3, the Random Forest and K-Nearest Neighbors

models are the top performers on this dataset, as they

consistently achieve high precision, recall, and F1

scores. The Logistic Regression and Support Vector

Machine models show weaker performance. At the

same time, the Artiﬁcial Neural Network model also

performs well, falling in between the top-performing

models and the weaker ones. While it’s noteworthy

that ANN forecasts rates for all potential outcomes

and offers real-time predictions, the instance-based

learning of KNN constrains its ability to provide real-

time predictions. Additionally, RF doesn’t furnish

rates for output classes, presenting results solely.

Figure 4 illustrates the initial ﬁve prediction out-

comes produced by the artiﬁcial neural network

(ANN) prediction engine using the Paris dataset. The

values displayed in the ﬁgure indicate the propor-

tions of status outputs, with each percentage corre-

sponding to a particular interpretation. This inter-

Table 1: Performance results (%) per model for the Belib

Dataset.

APPROACH RECALL PRECISION F1-SCORE

ARTIFICIAL NEURAL NETWORK 87.94 89.46 88.69

K-NEAREST NEIGHBOURS 94.13 94.10 94.11

LOGISTIC REGRESSION 72.16 66.65 60.44

RANDOM FOREST 95.56 95.54 95.41

SUPPORT VECTOR MACHINE 71.90 66.41 59.54

Figure 3: Performance results (%) per model for the Belib

Dataset.

pretation is further demonstrated in Figure 4, which

showcases the prediction outcomes of the model uti-

lizing ANN for the Belib dataset. In Figure 4a, the

charging spot is predicted to be 93.89% Available

(”Disponible”). In Figure 4b, the charging spot is

predicted to be 95.26% in maintenance (”En main-

tenance”) and 0.326% Available (”Disponible”). In

Figure 4c, the charging spot is predicted to be 79.79%

Available (”Disponible”) and 0.842% in the process

of commissioning (”En cours de mise en service”).

In Figure 4d, the charging spot is predicted to be

64.62% Available (”Disponible”) and 16.53% Busy

(”En charge”). In Figure 4e, the charging spot is pre-

dicted to be 99.92% Available (”Disponible”).

The ﬁndings from the application of machine

learning (ML) and deep learning (DL) approaches to

Estonian data are presented in Table 2. The K-Nearest

Neighbours model performs consistently across all

metrics. It achieves high precision, recall, and F1-

score, indicating its strong ability to classify instances

of different classes correctly. The Logistic Regression

model shows competitive performance. It achieves

high recall, indicating its effectiveness in identifying

positive cases, and maintains a balanced F1 score.

The Random Forest model demonstrates solid perfor-

mance across all metrics. It achieves high recall and

precision, resulting in a strong F1 score. The Support

Vector Machine (SVM) model performs well, with

high precision and recall values. The F1 score also

reﬂects a good balance between the two metrics. The

Artiﬁcial Neural Network (ANN) model achieves im-

pressive results, particularly in precision. Its recall

and F1-score are also high, indicating its effectiveness

in classiﬁcation tasks on this dataset.

Based on the results presented in Table 2 and Fig-

Accurate Recommendation of EV Charging Stations Driven by Availability Status Prediction

355

(a) First prediction outcome (b) Second prediction outcome (c) Third prediction outcome

(d) Fourth prediction outcome (e) Fifth prediction outcome

Figure 4: Results from the prediction model using ANN for the Belib database.

ure 5, all models exhibit strong performance on the

”Eneﬁt Volt Dataset”. The ANN, Logistic Regres-

sion, and SVM models consistently achieve high pre-

cision, recall, and F1 scores. The K-Nearest Neigh-

bours and Random Forest models also perform well

but have slightly lower F1 scores than the others.

Overall, these models appear to be well-suited for

classifying instances in the ”Eneﬁt Volt Dataset”.

Table 2: Eneﬁt Volt Dataset Performance Results (%) by

Model.

APPROACH RECALL PRECISION F1-SCORE

ARTIFICIAL NEURAL NETWORK 95.74 96.04 95.89

K-NEAREST NEIGHBOURS 94.19 94.11 94.21

LOGISTIC REGRESSION 96.63 94.35 94.81

RANDOM FOREST 96.02 94.78 95.17

SUPPORT VECTOR MACHINE 96.32 95.15 95.75

Figure 5: Eneﬁt Volt Dataset Performance Results (%) by

Model.

In Figure 6, there is a presentation of the initial

ﬁve prediction outcomes derived from applying the

artiﬁcial neural network (ANN) prediction engine to

the dataset from Estonia. The ﬁgures represented in

the visualization denote the percentages linked to dif-

ferent status outputs, as illustrated in Figure 6. This

ﬁgure illustrates the prediction outcomes of the model

utilizing ANN for the Eneﬁt Volt dataset. In Figure

6a, the charging spot is 99.69% available and free to

use. In Figure 6b, the charging spot is 70.46% in a

charging session and 29.46% available. In ﬁgure 6c,

the charging spot is 47.43% available; it is 35.7% in

a charging session and 16.83% occupied by someone

else. In Figure 6d, the charging spot is 99.56% avail-

able and free to use. In Figure 6e, the charging spot is

99.89% available and free to use.

The results of this paper demonstrate that the pre-

cision, recall, and F1 scores for all evaluation met-

rics surpass those of existing works in the literature

review. These ﬁndings are particularly noteworthy

given the differences in dataset features, which ac-

count for the varying evaluation scores of the models

for the Paris and Estonia datasets. The performance

results from both the ”Belib Dataset” and the ”Eneﬁt

Volt Dataset” demonstrate the effectiveness of vari-

ous machine learning models in tackling classiﬁcation

tasks. Notably, the Estonia dataset is deemed more

robust due to its inclusion of additional features such

as price per kWh and outlet types, contributing to its

higher overall scores than the Paris dataset.

Although the evaluation metrics for the Paris

dataset show that ANN has slightly lower scores than

RF and KNN, it should be emphasized that ANN

can provide real-time predictions, unlike KNN’s lazy-

learning approach. Overall, the models’ perfor-

mances indicate their capability to classify instances

accurately on both datasets. The choice of the best-

performing model would depend on the speciﬁc pri-

orities of the task at hand. Furthermore, the artiﬁ-

cial neural network (ANN) showcases outcomes in

the form of a percentage for each distinct category,

constituting a feature that is greatly favored. Fur-

thermore, given the potential for including more com-

plex features or a greater volume of records in future

datasets, the ANN model is likely to perform even

better with further improvements.

ICSOFT 2024 - 19th International Conference on Software Technologies

356

(a) First prediction outcome (b) Second prediction outcome (c) Third prediction outcome

(d) Fourth prediction outcome (e) Fifth prediction outcome

Figure 6: Results from the prediction model using ANN for Eneﬁt Volt dataset.

4 CONCLUSION

This paper’s initial focus was conducting a thorough

literature review and analyzing state-of-the-art re-

search on EV charging station availability prediction,

including its strengths, weaknesses, models, methods,

and datasets. By conducting a comprehensive review

of existing literature, a foundation and rationale have

been established to construct a scalable prediction en-

gine for real-time availability forecasting in electric

vehicle charging stations. The implementation pro-

cess involved scraping data from various sources, cre-

ating datasets, and applying feature engineering to the

data model. The paper then delved into using base-

line machine learning models on the pre-processed

dataset and subsequently building and training an arti-

ﬁcial neural network model, which served as the pre-

diction engine. In conclusion, the paper showcased

the results visually, and the performance of different

models was evaluated using standard metrics. The

intended outcome was realized through the develop-

ment of a prediction engine utilizing artiﬁcial neural

networks. This engine furnishes probabilities of elec-

tric vehicle charging station availability, showcasing

impressive evaluation scores that surpass the bench-

marks set by existing literature. There are several po-

tential avenues for future work to enhance the work

presented in this paper while also addressing the chal-

lenges posed by controlling the uncertainty of the dy-

namic arrival of EV charging requests.

• The manual nature of the data scraping and

feature engineering components suggests that

automation could yield further improvements.

Speciﬁcally, it may be possible to automatically

identify which features are most important and

which are less critical and to log differences that

arise when a particular feature is excluded. En-

gaging in this process additionally improves the

quality, precision, and dependability of the model.

• This study will also be extended to a larger scope,

which is a better understanding of capacity fade,

i.e., the battery’s ability to hold a charge dimin-

ishes over time. Beyond intrinsic features like

electrode degradation and solid-electrolyte inter-

face growth, we would also be eager to explore

the driving quality impact further.

• Last but not least, while the prediction engine is

currently limited to Paris and Estonia, it could be

readily adapted to other cities or countries.

ACKNOWLEDGMENTS

This work was partially funded by the HE project EN-

ERGETIC: https://energeticproject.eu/the-project/

under the reference number 101103667.

REFERENCES

Belib home. Available at: https://belib.paris/en/home.

City of Paris, Belib API. publicly Available.

data scraping. https://www.imperva.com/learn/application-

security/data-scraping/.

Eneﬁt VOLT. https://eneﬁtvolt.com/en/elektriauto-avalik-

laadimine.

Feature Engineering. https://www.javatpoint.com/feature-

engineering-for-machine-learning.

Almaghrebi, A., Aljuheshi, F., Rafaie, M., James, K., and

Alahmad, M. (2020). Data-driven charging demand

Accurate Recommendation of EV Charging Stations Driven by Availability Status Prediction

357

prediction at public charging stations using super-

vised machine learning regression methods. Energies,

13(6).

Balakrishnan, P. C. and Pillai, A. S. (2023). Design

and development of smart interoperable electric ve-

hicle supply equipment for electric mobility. Interna-

tional Journal of Electrical and Computer Engineer-

ing, 13(3):3509–3518.

El-Fedany, I., Kiouach, D., and Alaoui, R. (2023). A smart

system combining real and predicted data to recom-

mend an optimal electric vehicle charging station. In-

donesian Journal of Electrical Engineering and Com-

puter Science, 30(1):394–405.

El Halim, A., E. Bayoumi, W. El-Khattam, and A. M.

Ibrahim (2022). Electric vehicles: a review of their

components and technologies. International Journal

of Power Electronics and Drive Systems, 13(4):2041–

2061.

Engel, H., Hensley, R., Knupfer, S., and Sahdev, S. (2018).

Charging ahead: Electric-vehicle infrastructure de-

mand. (Exhibit 1):1–8.

George, V., Dixit, P., Dawnee, S., Agarwal, K.,

Venkataramu, V., and Giridhar, D. B. (2022). Com-

munication framework in an electric vehicle charging

station supporting solar energy management. Indone-

sian Journal of Electrical Engineering and Computer

Science, 28(1):49–57.

Giansoldati, M., Monte, A., and Scorrano, M. (2020). Bar-

riers to the adoption of electric cars: Evidence from

an Italian survey. Energy Policy, 146(July):1–10.

Gruosso, G., Mion, A., and Storti Gajani, G. (2020). Fore-

casting of electrical vehicle impact on infrastructure:

Markov chains model of charging stations occupation.

eTransportation, 6.

Hecht, C., Aghsaee, R., Schwinger, F., Figgener, J., Jarke,

M., and Sauer, D. U. (2022). Short-term prediction

of electric vehicle charging station availability using

cascaded machine learning models. 6th E-Mobility

Power System Integration Symposium and published

in the Symposium’s proceedings, (October).

Hecht, C., Figgener, J., and Sauer, D. U. (2021). Predicting

Electric Vehicle Charging Station Availability Using

Ensemble Machine Learning. Energies, (August):1–

24.

Hennlock, M. and WP4 Shift (2020). Strong link between

charging infrastructure and adoption of electric vehi-

cles. pages 1–4.

Intekin, F. (2022). Real-time availability prediction of elec-

tric vehicle charging spots. Master’s thesis, Tallinn

University of Technology. https://digikogu.taltech.ee/

en/Item/1fe83458-e52e-41da-a883-ee5fb1d78aad.

International Energy Agency (2023). Global EV Outlook

2023. Technical Report Geo.

Karike, S., Donepudi, S. R., and Raju, K. N. (2023). AC-DC

UPF charging circuit for two-wheeler electric vehicle

application. International Journal of Power Electron-

ics and Drive Systems, 14(1):11–24.

Kim, Y. and Kim, S. (2021). Forecasting charging demand

of electric vehicles using time-series models. Ener-

gies, 14(5).

Koohfar, S., Woldemariam, W., and Kumar, A. (2023). Per-

formance Comparison of Deep Learning Approaches

in Predicting EV Charging Demand. Sustainability

(Switzerland), 15(5).

Lee, Z. J., Li, T., and Low, S. H. (2019). ACN-Data: Anal-

ysis and applications of an open EV charging dataset.

In e-Energy 2019 - Proceedings of the 10th ACM Intl.

Conf. on Future Energy Systems, pages 139–149. As-

sociation for Computing Machinery, Inc.

Ma, T. Y. and Faye, S. (2022). Multistep electric vehi-

cle charging station occupancy prediction using hy-

brid LSTM neural networks. Energy, 244.

Qiao, F. and Lin, S. (2021). Data-driven prediction of ﬁne-

grained EV charging behaviors in public charging sta-

tions: Poster. e-Energy 2021 - Proceedings of the

2021 12th ACM Intl. Conf. on Future Energy Systems,

3466567(5):276–277.

Sao, A., Tempelmeier, N., and Demidova, E. (2021). Deep

Information Fusion for Electric Vehicle Charging Sta-

tion Occupancy Forecasting. IEEE Conference on In-

telligent Transportation Systems, Proceedings, ITSC,

2021-Septe:3328–3333.

Schulz, F. and Rode, J. (2022). Public charging infrastruc-

ture and electric vehicles in Norway. Energy Policy,

160.

Shahriar, S., A.R.Al-Ali, A.H.Osman, S.Dhou, and

M.Nijim (2021). Prediction of EV charging behav-

ior using machine learning. IEEE Access, 9:111576–

111586.

Soldan, F. e. a. (2021). Short-term forecast of electric vehi-

cle charging stations occupancy using big data stream-

ing analysis. 21st IEEE International Conference on

Environment and Electrical Engineering and 2021 5th

IEEE Industrial and Commercial Power System Eu-

rope, EEEIC / I and CPS Europe 2021 - Proceedings.

Tambunan, H. B., Sitanggang, R. B., Mafruddin, M. M.,

Prasetyawan, O., Kensianesi, Istiqomah, Cahyo, N.,

and Tanbar, F. (2023). Initial location selection of

electric vehicles charging infrastructure in urban city

through clustering algorithm. International Journal

of Electrical and Computer Engineering, 13(3):3266–

3280.

Yi, Z., Liu, X. C., Wei, R., Chen, X., and Dai, J. (2022).

Electric vehicle charging demand forecasting using

deep learning model. Journal of Intelligent Trans-

portation Systems: Technology, Planning, and Opera-

tions, 26(6):690–703.

Zhao, Y., Wang, Z., Shen, Z. J. M., and Sun, F. (2021).

Data-driven framework for large-scale prediction of

charging energy in electric vehicles. Applied Energy,

282(PB):116175.

Zhu, J., Yang, Z., Guo, Y., Zhang, J., and Yang, H. (2019a).

Short-term load forecasting for electric vehicle charg-

ing stations based on deep learning approaches. Ap-

plied Sciences (Switzerland), 9(9).

Zhu, J., Yang, Z., Mourshed, M., Guo, Y., Zhou, Y., Chang,

Y., Wei, Y., and Feng, S. (2019b). Electric vehicle

charging load forecasting: A comparative study of

deep learning approaches. Energies, 12(14):1–19.

ICSOFT 2024 - 19th International Conference on Software Technologies

358