A Novel Hybrid Neural Lyapunov Method With Low Conservatism for Power System Domain of Attraction Estimation

The strong representation ability of neural networks has provided a promising approach for power grid transient stability analysis. However, research in this area is still in the initial stage, and existing neural Lyapunov functions are characterized by several shortcomings, such as a high failure rate in falsification, difficulty in defining the validation region, and conservatism of the attraction domain. To address the above challenges, this article proposes a hybrid neural Lyapunov method with low conservatism. Unlike the existing neural Lyapunov functions, the proposed method is hybrid, combining linearized and nonlinear models. The region distant from the equilibrium point is analyzed using a neural Lyapunov function, while the region near the equilibrium point is analyzed using a linearized model. This hybrid strategy eliminates the need to strictly specify the initial validation region, mitigates the numerical stability problem of the neural network near the equilibrium point, and reduces the falsification region. Additionally, the improved trial-solution structure significantly simplifies the learning and falsification conditions and regularizes the optimization process to reduce the conservatism of the Lyapunov function. To further improve falsification efficiency, a domain of attraction (DOA) expansion algorithm with multithreaded chunked falsification is proposed. The method incrementally expands the validation region and has the potential to expand the isoenergetic surface defined by the LaSalle lemma. Applications of this method to transient stability analysis of synchronous generators and power electronics interfaced networked microgrids are presented, and multiple case studies are used to validate the suggested strategy. The results demonstrate that the method is generalizable and has a less conservative DOA estimate.