Data-driven energy demand forecasting for electric vehicle charging infrastructure

Abstract

In the era of Big Data and electric vehicles growth by market, data-driven methodologies assume a crucial role to create valuable information. The focus is on supporting the decision-making process for the development of an accurate charging infrastructure. Forecast analysis allows prediction of energy demand over the network. This supports growing trends with a consequent increase in customer satisfaction. By anticipating potential breakdowns due to infrastructure overloads, maintenance costs are reduced. In this paper, we focus on analyzing charging sessions data together with external data (weather and population information and energy/fuel prices) collected from diferent sources. The proposed methodology, named GEORGE (enerGy dEmand fOrecasting foR charGing infrastructurE), ofers a building-blocks based approach for the monthly energy demand forecasting. The approach is both generalisable and data-specific. We discuss the results of a classification learning approach to predict a belonging range of kwh for a charge point. In particular the most promising model has good performances in predicting high utilization and is more advantageous to support the company's decision-making process. Many possible developments are discussed to improve the prediction.

Keywords

• eol>Data Mining
• Electric Vehicle
• Predictive Model
• Applied Data Science
• Interpretable Model
• Energy Demand

Introduction

duce 2. Moreover, the overproduction of energy due to a disproportionate installation of charging stations In order to ensure transport emissions reduction in Eu- will decrease, [ 4 ]. A relevant player in this scenario is rope and to meet, for example, the Paris Agreement, [ 1 ], F2M eSolutions that designs innovative technologies to and Agenda 2030 objectives, [ 2 ], many goals need to be lead the transition to electric mobility. They ofer chargachieved over the next twenty to forty years. Two of the ing solutions and services that will make this change 17 goals, set by the Agenda, to be completed by 2030 at intuitive and seamless. Additionally, with the anticipaEuropean level are: the fight against climate change and tion of widespread adoption of Electric Vehicles (EVs) in the creation of sustainable communities and cities. These Europe, both public and private developers of charging objectives include the following targets. Integration of infrastructure heavily rely on proper data collection and national policies, strategies and plans against climate usage to make informed decisions, [ 5 ]. Electric mobility change and reduction of negative environmental impact is currently a topic of great interest for Europe and for of cities with regard to air quality. Transport sector in researchers. Many studies focus on how to take advangeneral and circulation of cars in particular directly in- tages from data to support Low-Carbon Road Transport lfuence greenhouse gas emissions, being responsible of Policies in Europe, [ 6 ]. Both the development of charging 12% of total emissions at European level, [ 3 ]. Thus, the infrastructures, [ 7 ] and the promotion of e-mobility to promotion of e-mobility through an interconnected and improve the user experience, [ 8 ] depend on an accurate optimized network of charging stations will help to re- utilization of data. Finding a general method to examine the impact of external factors on the energy demand Published in the Workshop Proceedings of the EDBT/ICDT 2023 Joint for charging infrastructure in Italy is currently an unreConference (March 28-March 31, 2023, Ioannina, Greece). solved topic in the literature. From a business side, this $ [e-mail] (S. Meddi); paper aims to answer to two main necessities. First, suptsaanraia.c.caevraqguliiotenl@[e-mail].(cTo.mCe(Srq.uCiatevlalig);lion); port and improve the management of charging stations [e-mail] (E. Manfredi); through a forecast of monthly energy demand. Second, [e-mail] (A. Regalia); let the client to identify targeted interventions in specific [e-mail] (R. Menolascino); areas. These actions can be supported by interpretable [e-mail] (G. Zardo)

© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License and accessible results that allow instant interaction with CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) data. To reach the goals we used F2M eSolutions’ charging session historical data integrated with external data many research orientate the analysis on data-centric apsources. Data collection and data preprocessing were proaches. the most time-consuming activities. The collection goes Studies cited in this section can be grouped in four from the company’s internal data (charge point ID, City, macro-categories: (i) application of data science to the Country, kwh, . . . ) to datasets containing information world of EVs, (ii) support for the transition to electric , (iii) about the population, weather data, gas oil/lpg/fuel and monitoring of infrastructure utilisation through Key Perenergy prices and characteristics of the territory, collected formance Indicators (KPIs) and (iv) study of the exogenous from third parties. Diferent data science models like factors that afect energy demand .

Random Forest, [ 9 ], and XGBoost, [ 10 ], have been trained Applications of data science methods to electric sceto determine the approach that better represents the phe- nario already include a wide spectrum of supervised nomenon under analysis. In the considered case study, and unsupervised learning approaches. Several works we followed a classification approach as it was the most have already been proposed in this context. [ 5 ] analyses consistent with business needs: predicting the kwh range the economic benefits of applying data-driven models, of monthly energy demand. In fact, unlike a daily fore- i.e. data mining, and machine learning techniques, in a cast, it provided a general overview of the infrastructure business context. Paper [11] provides a comprehensive status. This approach showed good results in predicting overview of use cases that link this scenario to data scithe observations belonging to the class with the high- ence, through machine learning algorithms. The authors est demand, which was the major focus from a business emphasise the scientific interest of the study in the field side. The final output is thus able to support business of e-mobility. A data-driven approach to extract useful decisions in order to achieve a more eficient charging information from electric vehicle charging events is also infrastructure usage. The use case on which the project suggested in paper [12]. In this case, a framework was has been developed is the 2022 Italian scenario. The pa- developed to characterise the demand for electric vehicle per is organized as follow. The first section is composed charging in a specific geographical area. A key step in by a brief literature review. It introduces the state-of-art this methodology is the one of data cleaning and foron the promotion and support of e-mobility through data. matting that is defined specifically for dealing with EVs Section two details the methodology, from the prepro- charging data. cessing step to the evaluation of the predictive results. Investing into realistic technological solutions, emSection three shows the results obtained by using the powering citizens and aligning action in key areas, i.e. proposed methodology on Italian use case. Finally, sec- industrial policy, ensure a fair transition to electric. The tion four summarizes the content of our work, providing authors of [13] and [ 8 ] focus their studies on predicting conclusions and suggestions for future improvements. charging point occupancy, to support this shift. In [13] the aim is to understand whether the city of San Diego has suficient charging points to meet the energy demand 1. Literature Review of EVs through a quantitative and qualitative analysis. In [ 8 ], the goal is to support users in planning their charging processes. In particular, they provide a double approach: classification to predict individual charging point occupancy and regression to predict overall charging station occupancy both in public and a workplace site. [14] comNowadays, collecting vast amounts of data has become a widespread practice in many scenarios. If done properly, it can be a powerful tool that can ofer significant beneifts to companies. With the growing market of electric vehicles, interconnection and data transfer technologies, pares performances of two diferent approaches. The 3. propose a general approach that can be expanded, ifrst one based the installation of new charging points thanks to the integration of external data sources, on a request by an electric vehicle driver. The second to investigate the influence of diferent factors on one chooses for the placement of a CP near strategic lo- future behaviors. cations, considering the decision of a local government.

Results show that not one rollout strategy is favorable over the other. Moreover, the best strategy needs to be 2. Methodology chosen according to municipal objectives, the maturity of the market and the technologies available. Here we present the GEORGE methodology, whose

The large amount of data collected brings with it the building-blocks representation is shown in Figure 1. It need to understand how to extract the most meaning- combines a solid theoretical background with a necessity ful knowledge. Monitoring properly constructed KPIs to solve real business needs. It uses historical data to usually allows to achieve the goal. [15] presents a study monitor charging stations and develop forecasting moddeveloped using data from the vast public charging infras- els to predict the monthly energy demand. It consists tructure of the Dutch metropolitan area. The researchers of a KDD process (Knowledge Discovery process from want to identify diferent charging patterns between five Data) adapted to business needs: support more eficient areas and relevant related KPIs. Thanks to forecasting development and utilization of the charging infrastrucand simulation models, they answer significant ques- ture. tions like where, when and what type of new charge GEORGE consists of four building blocks: points should be installed. In the development of the • data preprocessing step to clean and merge data study, [ 7 ], the authors propose a web-based dashboards from diferent sources; to explain particular well- or ill-performing charging • data transformation block to adapt the input data stations. The platform aims to support the projected structure to the final purpose; growth of electric mobility through the extraction of rel- • predictive analytics step to derive the most suitevant knowledge. Performances of the existing charging able descriptive model to perform accurate preinfrastructure, measured by KPIs developed, drive the dictions; know-how.

Many studies confirm that exogenous factors influence • tehvealguoaotidonneassndofint hteerpmroetdaetliotnhroofurgehsuslptsectoificamsseets-s charging behaviors. Just to name a few, [ 6 ] presents a rics. data processing platform TEMA (Transport tEchnology and Mobility Assessment) designed for supporting EU In the following sub-sections, we reported a detailed transport policies through big data. This study shows description of each building block. the implementation of a method capable of managing a significant amount of data from various sources. Thanks 2.1. Data Preprocessing to data-driven model, TEMA is able to recognize subtle connections and hidden patterns, performing customized Data preprocessing usually is the first step in the KDD analyses. Many governments started to base their charg- pipeline because prepares data for analysis in the most ing network definition on data-driven roll-out strategies. suitable way. For internal data, the json files recorded for The study [ 1 ] identifies and interprets the most impactful each charging sessions were stored in AWS (Amazon Web characteristics that are correlated with energy consump- Services) s3 bucket. Data was transformed into tables via tion. Authors ofer useful perspectives on what data need Athena, an analytics service of AWS, downloaded and to be utilized to create prediction models and to guide the saved locally as csv files. The external data was extracted planning and implementation of charging infrastructure. from open source data base in csv format. All input data

Literature shows that several prediction models enable was processed through python code using Anaconda relevant knowledge to be extracted from the data. Most environment (e.g. Spyder). GEORGE preprocessing was of the studies are very objective-specific and do not allow customized in three specific steps: (i) Data Ingestion and generalization. Main contributions brought by our study Cleaning, (ii) Data Integration from diferent sources and are: Alignment to the same granularity, (iii) Data Aggregation and Creation of new KPIs.

The objective of the Data Preprocessing was to have two macro categories of features:

1. bring the analysis to Italian level. The growing

circulation of electric vehicles in Italy allows to start analysing the Italian scenario. Until now this panorama is little mentioned in literature; 2. introduce a Sliding Window based approach to align the input structure to the final need;

Monthly Based Features (named Type 1 in the follow

ing) that are information available on a monthly basis and include: charging session information, weather data and energy, gas oil, fuel and lpg prices.

Social-Economic and Geographical Features (named

Type 2 in the following) refer to economic and social aspects of the population and to characteristics of the territory. This set includes, for example: average age of the population, employment rate and population density.

Since environmental factors can influence energy de

mand, we included external data in addition to charging behavior information. The comparison with domain experts led to the following considerations. External temperatures can afect performance of electric vehicles batteries and therefore the number of charging sessions. Moreover, energy prices have direct influence on the demand. On the other hand, since an electric car owner also owns a traditional vehicle, we decide to include gas oil, fuel and lpg prices. Social-economic and geographical factors are able to model the wealth of the population and therefore, indirectly, the ownership of electric vehicles.

Data ingestion for Type 1 features originated from various sources: KPIs. For each couple (, ) = (ℎ, ℎ), with at least one charging session, we computed the following metrics through a monthly level aggregation: total number of sessions, total session time, kwh provided and main statistics for weather values.

From them we computed the following additional indicators, average session time and charge point Occupation Rate: ave. sess. time [min] = tot. session time [min]

tot. #of sessions tot. session time [min] OR [%] = tot. #of days × 24 × 60[min] × 100

From Type 2 features prospective, the preprocessing

was easier. Data ingestion source included only Italian National Institute of Statistics (ISTAT) and data granularities were at Italian region and annual level. It was not possible to decrease the aggregation data level and therefore this set of information was simply added to the dataset, skipping step (ii) and (iii). 2.2. Data transformation • time level • geographical level and for this reason they had diferent granularities: • F2M eSolutions, for charge points and charging sessions data; • Meteo.it, for weather information; • Italian Government web site, for prices trends.

The step of data cleaning was managed in a specific way for each data source. In presence of internal anomaly, e.g., transmission or estimation error, the corresponding data was removed.

In case of external information not available, diferent strategies were implemented like replace the missing values with the closest geographical and temporal information.

During data integration, diferent data sources were aligned to the same granularity using supporting datasets, e.g., for geographical mapping.

Once, the collections had the same granularity we proceeded with the aggregation and definition of new

Data transformation step can be essential for the success

of the entire KDD process and it is typically very projectspecific.

In order to forecast energy consumption using historical data, we decided for a Sliding Window based approach.

This method uses a moving window to analyse sequential data and predict future values over a time period. The – state, region, city, longitude and latitude approach is justified by the capacity of historical trends for charge points and charging session; to afect current needs. We defined two variables, one to – region and province for weather informa- identify the time window width, X, and the other one to tion; define the time horizon for the prediction, or time forecast – state for prices; window, Y.

Data transformation process consisted of four steps to modify dataset structure, plus a preliminary step for the – year/month/day hour:minutes:seconds for encoding of all nominal variables.

charging sessions; Step 1 We divided dataset into smaller ones each with – year/month/day for weather information; information of historical data and prediction window, e.g. – month for prices. + months. months, then named as − , . . . , − 1, are for the construction of the time window while the last month, 0, refers to the time horizon.

Step 2 Not all charge points had data about the entire time window. In absence of useful information to reconstruct the lack of charging session data, we decided to ifll missing information with zeros. Thus, we modelled in the same way, a hypothetical failure of charge points and lack of charging sessions due to other reasons.

Step 3 Separately, each subdataset has been transformed to obtain the Sliding Window based structure.

Figure 2 the configuration with a time window of two months and a time horizon of one month.

Step 4 Finally, all sub-datasets were merged together, changes for each subset of charge points as shown in resulting in a single dataset where rows contained pre- Figure 2. dictors for + consecutive months. Then, through feature selection phase, conducted by correlation analysis, Recursive Feature Elimination with Cross Validation (RFECV ) and domain expert support, only the optimal set of predictors was identified to give rise to the best model. Usually, selecting a restricted set of top predictors helps to reduce noise in the model and makes result interpretation more straightforward and efective for business experts. 2.4. Evaluation and Interpretation

Model evaluation is an essential part to understand relia

bility of prediction against real value.

In this study, identifying charging points that had a growing trend and whose demand might overload the network was the focus of the analysis. Particularly, we 2.3. Predictive analytics were interested on avoiding misclassification of charge points that belonged to high demand class. Otherwise To obtain the data-driven model that best represented the this underestimation of the load might lead to potential analysed phenomenon, GEORGE required the validation risk of breakage in the infrastructure. In this sense, we of diferent algorithms. Therefore, the optimal model were primarily interested in reaching high values for was the one with the best combination of performance metrics of this class, named H in the following. and ability to provide results. To extract the best combination of model-purpose

Monthly energy forecasting could be approached using GEORGE integrated stratified-cross validation technique two diferent strategies: punctual or categorical value and, in case of limited data, the leave-one-out method. forecasting. While the first one is more specific, it would The latter technique consists of using a subset of data for require a large amount of data to obtain accurate models. training the model and one to validate it, applying then Therefore, a multi-class classification task would be of the learning process for each subset, [16]. For LOOCV wide interest on several case studies where, especially (Leave One Out Cross Validation) the test set consists of initially, data might be limited. In fact, from a business one observation. side was definitely more functional to have a range of For the classification task, chosen for business need, kwh rather than a point value. model performance was estimated based on misclassifi

We defined classes based on diferent levels of per- cation. Below the metrics selected to evaluate multi-class centiles, a statistical approach that grants balanced classes. classification models on class H, [17]: This procedure allowed the identification of diferent Precision amounts of load for the charging infrastructure.

SliGdiEnOgRWGiEndinowtegbraasteedd svtaruricotuusrefodreesccarsitbinedg amboodveel.s Museitnhg- = #oftootbasl.#coofrorebcst.layscsliagsnseifieddtoinclcalassss . ods chosen are tree-based: Random Forest (for details, Recall see [ 9 ]) and XGBoost (eXtreme Gradient Boosting), see [ 10 ]. These are two examples of ensemble learning techniques. In particular, Random Forest combines multiple #of obs. correctly classified in class decision tree models to improve general performance. = total #of observations belonging to class One of the advantages is that this technique has higher accuracy than single decision trees. Additionally it is F1 score for class H is the harmonic mean of the preinterpretable, robust to noise and outliers. Then, the sta- cision and recall tistical framework of XGBoost casts boosting as a numer- 2 * * ical optimization problem. The objective is to minimize 1 = + model loss by adding weak learners using a stochastic gradient descent-like procedure.

Models operated on historical data, corresponding to The interpretation of model performances allowed to predictors of Type 1 and Type 2 for − , . . . , − 1 re-evaluate some previous steps such as feature selection. and 0, to forecast the energy demand for 0, which

The trade of between results and characteristics of

models (such as interpretability) determined the choice of the final model to be applied to real time data.

In the following section we present the preliminary results of the optimal model for a use case in Italy.

3. Preliminary results

An accurate manipulation of the huge amount of data

that are available today is essential in order to make informed decisions. In this scenario the Energy Demand Forecaster for Italian charging infrastructure can assume a relevant role.

The goal was to create a model, based on temporal aggregation of input data, able to predict the monthly energy demand for single charge point. At the same time we were interested on understanding which factors assumed relevant role in the prediction and their interpretation.

The business value is to support informed decisions to help the network optimization. Moreover, adjusting the load of the charging infrastructure can help in avoiding overloads and breakages. 3.1. Data diferent widths for the time window and with the support of domain experts we set = 2 while for the forecast window = 1. This confirms that energy demand is linked to historical data by a short-term relationship.

The final dataset was thus composed by predictors for three consecutive months of all the charge points. Two months populated the data history while the kwh of the last one was identified as target variable. 3.3. Predictive analytics and evaluation

Since diferent algorithms were integrated into GEORGE

we performed various experiments, leveraging LOOCV, to compare performances.

We recall that, from a business side, it was more functional a classification approach that predicted the range of kwh for each charge point rather than a punctual value.

For each classification algorithm, we evaluated the number of classes, the best subset of features, the size of time window and the impact of hyperparameters. The best model-purpose combination was obtained for the Random Forest with 3 classes, 37 predictors summarized in Figure 3, a 2-month data history and the following hyperparameters: In this section we analyse the results of a case study n_estimators = 250 in Italy for which we had data collected from April to October 2022. On the territory the highest percentages max_depth = 109 were 20% in Piedmont and Lombardy and 10% in Lazio max_features = 18 and Emilia-Romagna. min_samples_leaf = 9

Dataset contained data such as spatial, temporal and min_samples_split = 7 energy with diferent granularities: state, region, city, longitude and latitude for the geographical level and year/- Where n_estimators is the number of trees in the formonth/day hour:minutes:seconds for charging sessions est, max_depth is the the maximum depth of each tree, for the time level. max_features is the maximum number of features to con

Principal information recorded in the internal data sider when splitting a node, min_samples_leaf is the the were: charging session duration, geographical localization, minimum number of samples required in a node to be at kwh provided during each sessions. Some charge points a leaf and min_samples_split is the minimum number of characteristics like if it was subject to any restrictions, i.e. observations in any given node in order to split it. was located in a parking station or it was not open 24/7, The model showed high precision, recall and F1 score were also included in the dataset. in the third (or H ) class, resulting thus more performing

On a monthly basis the mean of kwh provided was on forecasting high usage: about 14.000 kwh for a mean of 500 session per month. 3.2. Data Preprocessing

To obtain the best results from predictive model, we

had to understand and structure our data precisely. We started with data manipulations steps.

We exploited cases and limit setting, obtaining more than 3.000 charge points with at least one session and about 10.000 charging sessions. Through the aggregation and transformation steps we structured the dataset to develop the forecasting model. After the evaluation of improve the management of the charging network, but also the user experience of electric vehicle owners.

4. Discussion

Through GEORGE methodology, based on a solid theoretical background and motivated by concrete business needs, we succeeded in bringing tangible benefits to customer and company. In particular, it made possible the development of an interpretable and generalisable estimator for forecasting the energy demand of the charging infrastructure. The deployment of the model then enables the methodology to be applied to everyday scenario. The real time data stream is processed by GEORGE which extracts relevant patterns. The results are summarised in a customised dashboard that can monitor the state of the network and guide the operators’ interventions based on informed knowledge. Figure 5 summarizes key steps of the process on real-time data stream. This can not only

1Gini’s impurity index calculates each feature importance as the sum

over the number of splits (accross all tress) that include the feature, proportionately to the number of samples it splits.

For the specific use case, the biggest efort was on the

prediction of class with the highest values of kwh. Correct classifications allow to prevent potential breakages due to an overload of the infrastructure.

Although it is a preliminary work, it is promising because it shows good results for the prediction of the third class. In addition, thanks to the intuitive interpretation, results can be easily shown in an interactive dashboard which can display the history and the forecast for each charge point in navigable map.

Applying GEORGE to Italian’s charging infrastructure we can support business decisions with success. Among various benefits, the more relevant are: • network monitoring based on continuous collection of data; • promotion of underutilized areas to exploit the infrastructure; • load monitoring preventing potential breakages and consequent spread of moneys for maintenance; • implementation of prompt actions guided by predictive analysis; • support of growing trend in high demand areas, increasing customer satisfaction.

Although GEORGE is a general-purpose methodology,

the feature modeling step can not be fully automated. This phase must be modeled according to the final goals, with specific and constant support from domain experts. In addition, understanding which features afect the prediction is an essential aspect to target informed decisions, and it needs conscientious supervision.

GEORGE works with diferent type of data and this allows future developments of the analysis exploring aspects of the phenomenon as deep and specific as the business needs. It is possible to analyse the influence of population behaviors and characteristics. On the other hand, deeply understanding the impact of weather on battery performance and its consequence on charging demand assume an important role. In this sense, we can leverage GEORGE to address diferent use cases optimally, adjusting the blocks of the process based on new needs.

Together with the evolution of an electric reality, research progresses and so future directions of this study can be explored.

Possible improvements are: 1. develop an interactive dashboard to monitor the network and support informed decision through the visualization of both historical and forecast trends; 2. increase data history and quality. All the models that we applied are data driven so more quality data will definitely let the model to understand better the relationship between input and output. Decrease granularity, i.e., for the Type 2 predictors including information at municipality level, will help to explicate the dependence of energy demand on social-economics habits. Considering other type of external data may help to discover more hidden patterns, i.e., trafic information and electric vehicles growth by market for sure impact the forecast of energy demand. The first may support distribution of load through the infrastructure based on pick hours. The latter introduces dependencies on past and, possibly, future sales trend; 3. implement a hierarchical approach that includes a three-class classification model and a regression model for each class to also have a point prediction for each charge point.

In conclusion, improving the dataset is desirable to achieve a prediction based on a larger time forecast window, from a month to 3/6 months, to support the optimization of charging points installation plans.

Acknowledgments

This work was partially supported by

References

[1] M. Straka , R. Carvalho , G. V. D. Poel , L. Buzna , Analysis of energy consumption at slow charging infrastructure for electric vehicles , IEEE Access 9 ( 2021 ) 53885 - 53901 . doi: 10 .1109/ACCESS. 2021 . 3071180 .

[2] A. per la coesione territoriale , Agenda 2030 per lo sviluppo sostenibile , 2015 . URL: https: //www.agenziacoesione.gov.it/comunicazione/ agenda-2030 - per -lo-sviluppo-sostenibile/.

[3] C. A. European Union , Co2 emission performance standards for cars and vans, 2020 . URL: https://climate.ec. europa.eu/eu-action/ transport-emissions/ road-transport-reducing-co2-emissions-vehicles/ co2-emission-performance-standards-cars-and-vans_ en.

[4] U. N. E. Programme , Supporting the global shift to electric mobility, 2021 . URL: https://www.unep.org/explore-topics/ transport/what -we-do/electric-mobility/ supporting-global-shift-electric-mobility.

[5] J. Sung , M. Troilo , N. Howarth , Better energy eficiency policy with digital tools , 2021 . URL: https://www.iea.org/articles/ better -energy-eficiency-policy-with-digital-tools.

[6] M. De Gennaro , E. Pafumi, G. Martini, Big data for supporting low-carbon road transport policies in europe: Applications, challenges and opportunities , Big Data Research 6 ( 2016 ) 11 - 25 . URL: https: //doi.org/10.1016/j.bdr. 2016 . 04 .003. doi: 10 .1016/ j.bdr. 2016 . 04 .003.

[7] S. Maase , X. Dilrosun , M. Kooi , R. Van den Hoed, Performance of electric vehicle charging infrastructure: Development of an assessment platform based on charging data , World Electric Vehicle Journal 9 ( 2018 ). URL: https://www.mdpi.com/2032-6653/9/2/ 25. doi: 10 .3390/wevj9020025.

[8] A. Ostermann , Y. Fabel , K. Ouan , H. Koo , Forecasting charging point occupancy using supervised learning algorithms , Energies ( 2022 ). URL: https://doi.org/10.3390/en15093409. doi: 10 .3390/ en15093409.

[9] T. Hastie , R. Tibshirani , J. Friedman, The Elements of Statistical Learning . Data Mining, Inference, and Prediction , Springer, New York, NY, 2009 .

[10] T. Chen , C. Guestrin , Xgboost: A scalable tree boosting system , in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , KDD '16, Association for Computing Machinery, New York, NY, USA, 2016 , p. 785 - 794 . URL: https://doi.org/ 10.1145/2939672.2939785. doi: 10 .1145/2939672. 2939785.