Deep Learning-Based Multiclass Attacks Detection and Classification in Automatic Generation Control

This research employs the ResNet-50 deep learning model to accurately detect and classify four types of false data injection (FDI) attacks (pulse, random, ramp, and scaling) in the automatic generation control (AGC) power system. Leveraging original signal data, measuring frequency, and tie line power deviations, the approach utilizes continuous wavelet transform and scalogram for input signal processing, enabling autonomous feature learning during training. To tackle multiclass classification challenges, error-correcting output codes with a dense random design coding method are applied. Rigorous evaluation involves gentle adaptive boosting (Gentle AdaBoost), support vector machine and its variant model, and k-nearest neighbors as binary learners, with gentle boost achieving a notable 100% accuracy in distinguishing and classifying all FDI attacks. This precise classification is crucial for proposing effective mitigation strategies, enhancing AGC cybersecurity in smart grids.