Machine learning-based modeling of usage trends and demand forecasting in electric vehicle charging stations

With the increasing adoption of electric vehicles, demand forecasting for charging infrastructure has gained critical importance in terms of energy supply security and power grid management. In this context, a comprehensive comparative analysis was conducted using four datasets constructed at different temporal resolutions. The first dataset consists of raw hourly measurements, the second includes daily aggregated values, the third comprises weekly totals, and the fourth represents a 24-hour ahead forecasting structure based on an hourly sliding window format. Each dataset was modeled using a wide range of forecasting approaches, including classical time series models such as ARIMA (Autoregressive Integrated Moving Average) and SARIMA (Seasonal Autoregressive Integrated Moving Average), parametric Holt–Winters exponential smoothing methods, machine learning-based algorithms such as Support Vector Regression (SVR), Random Forest, XGBoost, LightGBM, Gradient Boosting Regressor (GBR), and Prophet, as well as deep learning architectures including Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM) networks. Data preprocessing procedures—including missing value imputation, feature scaling, and seasonal component decomposition—were carefully implemented. Additionally, hyperparameter optimization for each modeling approach was conducted experimentally to ensure fair and robust performance evaluation. The analysis results demonstrate that short-term demand fluctuations are most effectively captured by LSTM-based deep learning models and the Prophet framework. For the first and third datasets, two-layer LSTM architectures achieved the lowest error metrics, while Prophet, employing a multiplicative seasonality structure, provided the highest forecasting accuracy for the second dataset. In the case of the fourth dataset, which focuses on hourly-ahead demand prediction, the BiLSTM architecture—with a single-layer configuration and low regularization settings—exhibited the best generalization performance. Although classical ARIMA and SARIMA models produced balanced outcomes with respect to information criteria, they were less effective in minimizing forecasting errors compared to deep learning and Prophet-based approaches. Similarly, the contribution of parametric exponential smoothing methods remained limited. In contrast, machine learning- xv based models such as XGBoost and LightGBM may be preferable in scenarios where low computational cost and interpretability are prioritized. In conclusion, LSTM–BiLSTM architectures and Prophet are recommended for short-term planning and real-time optimization tasks in electric vehicle charging demand forecasting. Future studies may further enhance predictive performance by integrating exogenous variables and exploring hybrid modeling frameworks.