Quantum-Inspired Graph Attention and Hybrid Deep-Learning Ensemble for Electric-Vehicle Charging Energy Prediction

Abstract Currently, electric vehicles (EVs) are gaining widespread attention in various countries. The accurate prediction of charging station load is essential for optimized smart grid operations. The improper prediction results lead to various issues like peak load stress and grid congestion. Existing prediction models fail to handle low spatial resolution and spatial heterogeneity of charging data. In this work, an ensemble deep-learning model is proposed for the accurate prediction of charging load. This model integrates a quantum-inspired graph attention network (QGNATNet) with a hybrid model combining long short-term memory (LSTM), static graph convolutional embeddings, and gated residual networks. QGNATNet uses quantum-inspired encoding of time-series charging sessions to capture fine-grained spatiotemporal correlations using attention-driven message passing across graph nodes. The hybrid model learns sequential temporal dependencies through LSTM layers. In addition, it incorporates contextual embeddings from a static GCN and enhances feature representations using gated residual blocks and multihead attention inspired by temporal fusion transformers (TFT). The outputs of both models are combined via ensemble averaging to improve robustness and prediction accuracy. Experimental evaluation on four benchmark data sets—Oslo and Trondheim in Norway, Palo Alto, California, and Dundee, Scotland—demonstrated that the ensemble model consistently outperformed state-of-the-art approaches. The ensemble model achieved a root-mean squared error (RMSE) of 16.57, mean absolute error (MAE) of 11.64, mean absolute percentage error (MAPE) of 48.36%, and a coefficient of determination ( R 2 ) of 0.96.