Physical Mechanism Enabled Neural Network for Power System Dynamic Security Assessment

Data-driven artificial intelligence technologies have emerged as increasingly fascinating tools for assessing power system security. However, their inherent mechanism of inexplicability and unreliability now limits their scalability in power systems. To address this problem, this paper proposes a neural network design method empowered by physical mechanisms for power system security assessment. It incorporates geometric characteristics of dynamic security regions into the network training process and constructs connections between network structure and system's unstable mode, which can perform security assessment with a neural network efficiently while ensuring physical plausibility. Furthermore, a credibility evaluation mechanism is established to ensure credibility of neural network predictions and reduce misclassifications. Finally, effectiveness of the proposed method is verified on test systems. Methods and considerations in building a neural network with interpretable structures and credible predictions can provide a reference for machine intelligence applied in other industrial systems.