Generative Self-Supervised Learning for Cyberattack-Resilient EV Charging Demand Forecasting

The Electric Vehicle (EV) market is experiencing unprecedented growth. Accurate prediction of EV charging demand is essential for transportation system operations, such as real-time traffic management, route optimization, and station utilization planning. However, Cyber threats can compromise the accuracy of charging demand predictions, leading to significant disruptions in transportation services, e.g., suboptimal station management, unexpected congestion at charging facilities, and degraded service quality for EV users. This study introduces Generative Multi-task Self-supervised Learning for Prediction (GenS2-P), a cyberattack-resilient framework designed to ensure reliable charging demand predictions under adversarial conditions. GenS2-P incorporates a Denoising/Reconstruction AutoEncoder (DRAE) and a spatio-temporal prediction model to tackle the dual challenges of data poisoning and DoS attacks. By leveraging generative self-supervised learning and multi-task learning, GenS2-P effectively extracts spatio-temporal patterns to denoise and reconstruct data corrupted by cyberattacks. Experimental evaluations using real-world EV charging data demonstrate that GenS2-P significantly reduces prediction errors and mitigates cyberattack-induced disruptions. This improved prediction reliability enables more effective charging infrastructure management and supports robust transportation system operations even under adverse conditions.