FEMUS-Nowcast: A Robust Deep Learning Model for Sky Image–Based Short-Term Solar Forecasting Under Adversarial Attacks

Accurate short-term solar power forecasting (nowcasting) facilitated by smart devices and cyberinfrastructure, which uses sky images and artificial intelligence (AI)–based models, is susceptible to cyberattacks. This study investigates the vulnerabilities of deep learning (DL) and artificial neural network (ANN)–based sky image–based nowcasting models to adversarial attacks such as fast gradient sign method (FGSM), projected gradient descent (PGD), and a mixed attack template, along with proposing a feature extraction–based multi-unit solar (FEMUS)-Nowcast model. Results reveal that adversarial attacks significantly degrade all models’ accuracy and lead them to an unusable state. Moreover, FGSM is found to be the most severe attack, with root mean square error (RMSE) increasing by 5–16 times and mean absolute error (MAE) increasing by 4–12 times compared to the normal scenario under maximum perturbation. As the proposed FEMUS-Nowcast outperforms models of existing literature, reducing RMSE by 48% and 25% under normal conditions, adversarial training is adapted to enhance its robustness in the presence of cyberattacks. Furthermore, adversarially trained (AT) FEMUS-Nowcast shows no RMSE or MAE trade-offs under all scenarios. Additionally, the AT FEMUS-Nowcast model demonstrates high resilience against advanced attacks, including iterative FGSM (I-FGSM) and momentum I-FGSM (MI-FGSM), confirming its reliability and robustness across diverse attack scenarios.