Skip to main navigation Skip to search Skip to main content

Machine Learning zur Prognose von Abrechnungsverlusten bei Elektrofahrzeug-Ladevorgängen
: Binärer Klassifikationsansatz und Modellevaluierung

  • Bernadette Ackerl

    Student thesis: Master's Thesis

    Abstract

    With the increasing adoption of electric mobility, e-Mobility Service Providers (eMSPs) face financial challenges due to unsuccessfully billed charging sessions. Billing losses occur when electric vehicle drivers (EV-Drivers) charge, but the payment fails to reach the eMSP for various reasons. This thesis investigates, for the first time, the applicability of machine learning for the preventive prediction of such billing losses in charging processes, thereby addressing a research gap at the intersection of machine learning and electric mobility. The objective of this work is to develop a model capable of predicting, before the start of a charging session, whether billing will be successful. Two research questions guide the investigation: which machine learning algorithms are best suited for this prediction, and which features most strongly influence the model’s decision? The problem is modeled as a binary classification problem with the classes billed and not billed. The investigation is based on a dataset of 151.210 charging sessions from a German company covering the period from July 2023 to May 2024. In total 14.637 sessions, almost 10%, could not be billed. Implementation is conducted using five different supervised learning algorithms: Random Forest Classifier, XGBoost Classifier, Multi-Layer Perceptron, Logistic Regression, and Support Vector Classifier. Among these, the Random Forest Classifier performs best on the test dataset containing 21.056 charging sessions, including 1.626 that were not billed. The Balanced Accuracy is 0.91, corresponding to 91% correct classification with balanced weighting of both classes. The F1-Score is 0.89 as the harmonic mean between precision and recall. The ROC-AUC reaches 0.98, demonstrating a very high ability of 98% for correct class discrimination. Feature importance analysis reveals that historical data of the EV-Driver represents the strongest predictor. In practice, the model is applicable for contract-based payments, as the required historical data of an EV-Driver is available there. This thesis demonstrates that data-driven approaches are not only technically feasible but also economically valuable for eMSPs. They provide a scalable opportunity to significantly reduce financial losses from billing risks in electric mobility.
    Date of Award2025
    Original languageGerman (Austria)
    Awarding Institution
    • FH OÖ
    SupervisorGerald Ostermayer (Supervisor)

    Studyprogram

    • Data Science and Engineering

    Cite this

    '