Load Frequency Control of Smart Grid Based on Data-Driven FDI Attacks Detection and Data Repair

This paper studies load frequency control (LFC) of smart grids under false data injection (FDI) attacks and proposes a data-driven self-healing scheme for smart grids, which integrates FDI attack detection, data repair, and LFC into a unified framework. An autoencoder is utilized to perform dimensionality reduction and feature extraction, in which a multi-head self-attention mechanism is introduced to improve the detection of FDI attack characteristics. Additionally, the autoencoder is embedded within an advanced generative adversarial network framework to successfully detect FDI attacks by differentiating between anomalous and normal data. Given the high labeling costs and scarcity of labeled data in power system datasets, this approach requires only a small amount of labeled data and a large amount of unlabeled data for training. The FDI attack detection results are applied to a multi-agent reinforcement learning asynchronous advantage actor-critic algorithm for realtime data repair and security control, where centralized training and decentralized execution are combined to enhance the smart grids adaptability and overall stability under FDI attacks. A shared convolutional neural network is employed for centralized training to maximize global rewards and ensure overall system performance, while decentralized execution enables each agent to independently access grid data and attack detection results within their respective regions, autonomously performing data repair and power generation adjustments to control load frequency. Furthermore, a cooperative reward function is built to reduce data repair errors and area control error across regions. Validation of this scheme on the IEEE 14-bus system demonstrates that it not only effectively detects FDI attacks but also achieves data repair and secure control.