Electric vehicles charging stations load forecasting based on hybrid XGBoost-BiLSTM model

Area of Science:

• Energy Systems
• Artificial Intelligence
• Electrical Engineering

Background:

• Growing electric vehicle (EV) adoption necessitates accurate load forecasting for Electric Vehicle Charging Stations (EVCS) to ensure grid stability and optimize energy management.
• Short-term, station-level load prediction is critical for operational efficiency in managing EVCS.
• Existing forecasting models face challenges in capturing the complex temporal dynamics of EV charging patterns.

Purpose of the Study:

• To develop and evaluate a hybrid XGBoost-BiLSTM stacking model (Hybrid 3) for short-term, hourly EVCS load forecasting.
• To benchmark the proposed Hybrid 3 model against various statistical, machine learning, deep learning, and ensemble methods.
• To analyze the robustness, generalization capabilities, and key predictors for EVCS load forecasting.

Main Methods:

• Utilized the Adaptive Charging Network (ACN-Caltech) dataset, preprocessing 31,424 raw charging sessions into 14,496 cleaned sessions.
• Implemented a stacking ensemble model (Hybrid 3) combining XGBoost and BiLSTM with an XGBoost meta-learner.
• Conducted performance evaluation using Mean Absolute Error (MAE) and R-squared (R²) metrics, complemented by five-fold walk-forward validation and cross-site evaluation on synthetic data.

Main Results:

• The Hybrid 3 model achieved an MAE of 2.6870 kWh and R² of 0.6395 on cleaned data, outperforming standalone BiLSTM by 3.4%.
• While slightly underperforming the top gradient-boosting ensemble, Hybrid 3 demonstrated robustness with a mean MAE of 2.5351 kWh across validation folds.
• Feature importance analysis revealed log-transformed charging duration as the most significant predictor, indicating the importance of user behavior.

Conclusions:

• The Hybrid 3 model offers an effective balance between predictive accuracy and model complexity for EVCS load forecasting.
• While Hybrid 3 is competitive, gradient-boosting ensembles are recommended for scalable, real-time applications.
• Site-specific temporal patterns significantly impact forecasting generalization, suggesting the need for localized model tuning.