Exploring the potential of machine learning methods for predicting charging power curves

Tools

Jovanovic, Raka and Bayhan, Sertac and Bayram, I. Safak; (2025) Exploring the potential of machine learning methods for predicting charging power curves. In: 2025 10th IEEE Workshop on the Electronic Grid (eGRID). IEEE Electronic Power Grid (eGrid) . IEEE, GBR. ISBN 9798331593643 (https://doi.org/10.1109/eGRID63452.2025.11255577)

[thumbnail of Jovanovic-etal-IEEE-eGRID-2025-Exploring-the-potential-of-machine-learning-methods-for-predicting-charging-power-curves]
Preview
 Text. Filename: Jovanovic-etal-IEEE-eGRID-2025-Exploring-the-potential-of-machine-learning-methods-for-predicting-charging-power-curves.pdf
Accepted Author Manuscript
License: Creative Commons Attribution 4.0 logo
Download (877kB)| Preview

Abstract

This study addresses the challenge of predicting the charging power of electric vehicles (EVs) during an entire charging session using weather data. Data were collected from a Aiways U5 with a 60kWh battery across approximately 120 charging sessions between May 2023 and January 2025, resulting in over 55,000 individual time-stamped records. The prediction task was formulated as a regression problem, where the goal is to estimate the charging power at each time step based on the current state-of-charge (SoC), elapsed time since the session start, and external weather conditions such as temperature, humidity, wind speed, and cloud cover. Multiple machine learning models were evaluated, including linear regression (LR), random forest (RF), support vector regression (SVR), gradient boosting regression (GRB), and a feedforward neural network (NN). The performance was assessed using mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R2). Results indicate that GRB and RF outperform other models, with GRB achieving the best RMSE and R2 scores, while RF slightly outperforms in MAE. The neural network underperformed compared to tree-based methods, likely due to the limited diversity of training sessions despite the large number of records. In addition, a detailed analysis of the characteristics and distribution of the prediction errors is presented.

ORCID iDs

Jovanovic, Raka, Bayhan, Sertac and Bayram, I. Safak ORCID logoORCID: https://orcid.org/0000-0001-8130-5583;
CORE (COnnecting REpositories)