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

## TL;DR

This paper proposes a hybrid XGBoost-BiLSTM model for forecasting electric vehicle charging station loads, achieving improved accuracy compared to standalone models.

## Contribution

The novel contribution is the development and evaluation of a hybrid XGBoost-BiLSTM stacking model for EVCS load forecasting.

## Key findings

- Hybrid 3 achieved an MAE of 2.6870 kWh and R2 = 0.6395, showing a 3.4% improvement over standalone BiLSTM.
- Gradient-boosting ensembles outperformed Hybrid 3 in scalability and real-time forecasting.
- Log-transformed charging duration was identified as the dominant predictor in feature importance analysis.

## Abstract

Accurate load forecasting for Electric Vehicle Charging Stations (EVCS) is critical for optimizing energy management and ensuring grid stability amid growing electric vehicle adoption. This study investigates short-term, hourly load forecasting at the station level using a hybrid XGBoost–BiLSTM stacking model (Hybrid 3) with an XGBoost meta-learner. From the Adaptive Charging Network (ACN–Caltech) dataset (April 25, 2018–September 13, 2021), 31,424 raw charging sessions were preprocessed, yielding 14,496 cleaned sessions for modeling. These were split into 80% training and 20% testing sets using a fixed random seed (42) for reproducibility. Hybrid 3 was benchmarked against 24 alternative models spanning statistical (e.g., Persistence, SARIMAX), machine learning (e.g., XGBoost, LightGBM), deep learning (e.g., BiLSTM, CNN), and ensemble methods. On cleaned data, Hybrid 3 achieved an MAE of 2.6870 kWh and R2 = 0.6395—a 3.4% improvement over standalone BiLSTM—while slightly underperforming the top Boosting ensemble (XGBoost + BiLSTM + LightGBM). Robustness was confirmed via five-fold walk-forward validation (mean MAE = 2.5351 kWh, SD = 1.2885). Cross-site evaluation on an independent synthetic dataset (n = 1,965,239 sessions) showed reduced generalization, highlighting site-specific temporal patterns. One-Way ANOVA (p = 0.2073, η2 = 0.2062) indicated no statistically significant but practically relevant differences among top models. Feature importance analysis identified log-transformed charging duration (importance = 0.376) as the dominant predictor, aligning with real-world EV behavior. Overall, Hybrid 3 balances accuracy and complexity effectively, though gradient-boosting ensembles remain preferable for scalable, real-time EVCS forecasting.