A novel hybrid deep learning model for electric vehicles load forecasting under cyber-attacks

Accurate short-term electric vehicle (EV) load forecasting is essential for effective energy management and reliable operation of smart grids. However, the growing prevalence of cyber-attacks, particularly false data injection attacks (FDIAs), poses significant challenges to the integrity and reliability of forecasting models. This study proposes a novel hybrid framework, termed GAN-CLSTM, which integrates Generative Adversarial Networks (GANs) with Convolutional Long Short-Term Memory (CLSTM) networks to enhance forecasting accuracy while ensuring robustness against data integrity attacks. The GAN component is employed to preprocess input data, effectively denoising and mitigating anomalies, while the CLSTM module captures temporal dependencies and nonlinear relationships in EV charging behavior. The proposed model was validated using real-world data from a public EV charging station at the Caltech campus, USA. Two experimental scenarios were considered: (i) load forecasting under clean data conditions, and (ii) forecasting under scaling-based FDIA attacks. Comparative analysis against benchmark models including CNN, BiLSTM, CNN-LSTM, and GAN-LSTM demonstrated the superiority of the proposed approach. Under normal conditions, GAN-CLSTM achieved an R-value of 99.70% with the lowest error indices, while under attack scenarios, it maintained high resilience with an R-value of 98.84%, outperforming all competing models. The results confirm that GAN-CLSTM offers a reliable and cyber-secure forecasting solution, maintaining reliable forecasting performance and enhanced resilience under the considered adversarial conditions, thus contributing to the resilience of next-generation smart grids.