Secure electric vehicle charging station demand forecasting under adversarial and false data injection attacks

Forecasting the charging demand of electric vehicle charging stations (EVCSs) is critical for urban planning, resource allocation, policy development, and efficient grid management. However, reliance on sensor-collected data transmitted through various communication channels poses significant cybersecurity risks to forecasting models. This study evaluates the vulnerability of several commonly used forecasting models named random forest (RF), convolutional neural network (CNN), convolutional long short-term memory (ConvLSTM), and bidirectional convolutional long short-term memory (BiConvLSTM), under simulated cyber-attacks. Different attack scenarios, including fast gradient sign method (FGSM) and basic iterative method (BIM)-based adversarial attacks and scaling-based false data injection (FDI) attacks, are considered with varying attack volumes and perturbations. Metrics, such as root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), are employed to assess and compare the accuracies of different models. Results indicate that forecasting models exhibit significant performance degradation under these cyber-attacks. As a countermeasure, adversarial training has been employed to mitigate the impact of such attacks and has proven to be highly effective. The experimental observations in this research elucidate the impact of cyber-attacks along with a defense mechanism to mitigate economic and technical risks to the EVCS, fostering the future development of accurate and cyber-resilient forecasting methodologies essential for advancing both academic and industrial domains of EVCS demand forecasting.