A hybrid deep learning framework for multivariate energy forecasting and peak load prediction in electric vehicle charging infrastructure

Electric vehicles (EVs) play a pivotal role in sustainable mobility and the decarbonization of transport sector. Accurate forecasting of energy demand at EV charging station clusters (EVCSCs) is essential for optimizing smart grid operations, reducing peak load stress, and guiding cost-effective infrastructure deployment. This study presents a unified and scalable hybrid deep learning framework that integrates convolutional neural networks (CNNs), bidirectional LSTM and GRU units (BiLSTM/BiGRU), and attention mechanisms to model complex spatiotemporal dynamics in EV charging data. Domain-specific feature engineering, incorporating lag and rolling window statistics, is combined with Keras Tuner-based hyperparameter tuning and Genetic Algorithm-based time-step optimization to enhance the models’ adaptability and performance across diverse urban environments. Unlike conventional univariate methods, our framework supports multivariate forecasting by jointly predicting daily energy consumption, greenhouse gas (GHG) savings, and gasoline displacement. It also includes weekly peak day forecasting, enabling utility operators to anticipate high-load periods and manage demand-side resources more effectively. The models are trained and evaluated on four real-world public datasets (Palo Alto, Boulder, Dundee, Perth) representing varying consumption patterns and feature granularities. The results demonstrate superior accuracy in terms of RMSE, MAE, and R2 (up to 0.99), consistently outperforming baseline and state-of-the-art approaches. By integrating operational forecasting with environmental impact metrics, this framework provides a scalable and data-driven tool for intelligent EV energy management system (EMS). It contributes to sustainable energy planning and supports decarbonization strategies aligned with global Net Zero and sustainable development goal (SDG) targets.