Research on load frequency control system attack detection method based on multi-model fusion

Load frequency control (LFC) in power systems faces increasingly complex cyber-physical attack threats, while existing detection methods have limited capability to identify intelligent attacks. This paper constructs an LFC system model considering dynamic response characteristics and establishes a reinforcement learning-based method for generating multiple attack strategies, covering typical scenarios such as false data injection (FDI) and load switching attacks. A multi-model fusion attack detection framework is proposed, integrating (Long Short-Term Memory) LSTM supervised learning and autoencoder unsupervised learning algorithms, with an adaptive weight adjustment mechanism that dynamically optimizes detection strategies. Experimental results demonstrate that the fusion mechanism achieves 99.4% comprehensive identification accuracy across four system states, outperforming single supervised models (98%) and single unsupervised models (76.4%). Detection accuracy exceeds 99% for three different frequency characteristic attacks, with an average detection delay of only 0.12 seconds. The fusion mechanism effectively reduces false positive and false negative rates (FNRs), showing significant advantages in identifying and defending against unknown attacks, providing a new approach for LFC system security protection.