Detection of Cyber and Physical Disturbances in Power Systems Using Data-Driven Deep Learning Methods

This research investigates the use of advanced deep learning architectures to identify cyber incidents and physical disturbances within smart grid systems. By employing real-time data acquired from Phasor Measurement Units (PMUs), the research evaluates the performance of cutting-edge models in detecting and classifying anomalies in electrical networks. Specifically, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks are implemented to recognize a diverse range of cyber–physical disruptions, offering insights into their effectiveness in maintaining grid reliability and resilience. To enhance detection accuracy, a transfer learning model is proposed by integrating the complementary strengths of GRU architecture. This model outperforms conventional deep learning approaches in terms of classification performance. The deep learning model is trained on a dataset that includes instances of False Data Injection (FDI) attacks under both AC and DC power flow conditions, as well as Denial of Service (DoS) attacks, all of which are part of the cyber-attack scenarios. These are coupled with physical events, including line outages, generator outages, faults, and other operational contingencies. Extensive case studies are conducted on the IEEE 14-bus, 24-bus, 39-bus, and 118-bus systems, with all simulations implemented in Python. The results indicate that the proposed transfer learning approach achieved a high detection accuracy of up to 99.85%, along with lower Root Mean Square values, demonstrating its effectiveness in identifying cyber–physical events. The proposed transfer learning framework achieves consistent performance gains across different scenarios, with a maximum improvement of 0.32% over the benchmark GRU model. The presented method was compared again against traditional machine learning algorithms SVM, RFC and other deep learning models (Autoencoder, 1D-CNN, and Transformer). The comparative results confirm the superior performance and robustness of the proposed framework for cyber–physical event detection in power grids.