Data Recovery for False Data Injection Attacks of Power Systems Based on Lightweight Variational Convolution Auto-Encoder

Power system state estimation (PSSE) plays a vital role in stable operation of modern smart grids, while it is vulnerable to cyber attacks. False data injection attacks (FDIA), one of the most common cyber attacks, can tamper with measurement data and bypass the bad data detection (BDD) mechanism, leading to incorrect PSSE results. This paper proposes a data recovery framework for FDIA based on variational convolution auto-encoder (VCAE) to ensure continuous monitoring of PSSE. VCAE combines deep learning ideas with Bayesian inference. Besides, VCAE uses convolution and deconvolution operations with excellent feature capture capabilities in the encoder and decoder network, respectively, to effectively restore the abnormal values after FDIA to the values close to normal operation. Moreover, knowledge distillation (KD) is used to compress the VCAE model, making it possible to deploy a lightweight model on equipment with limited resources. Case studies are undertaken on IEEE 14-bus system under different attack intensities and degrees to evaluate the recovery performance of VCAE. Simulation results show that the mean absolute error (MAE) and mean absolute percentage error (MAPE) of VCAE are lower than the comparative generative models. Moreover, the satisfactory recovery performance of IEEE 30-bus and 118-bus systems verifies the scalability of the proposed model. In addition, KD allows the lightweight VCAE scale to reach about one-tenth of the original VCAE scale with almost no increase in MAE and MAPE.