Deep Lyapunov Learning: Embedding the Lyapunov Stability Theory in Interpretable Neural Networks for Transient Stability Assessment
Jiacheng Liu, Jun Liu, Rudai Yan, Tao Ding
Abstract
The machine learning-based transient stability assessment (TSA) has shown satisfactory accuracy while been limited by the lack of interpretability. This letter thereby presents a novel deep learning paradigm that naturally embeds the Lyapunov stability theory of dynamic systems, in which approximating Lyapunov functions (LFs) is transformed into traditional regression or classification tasks. The Lyapunov stability theory is firstly extended and then integrated into a specific neural network structure, which consists of a flexible LF approximator and its corresponding gradient adjoint network. It is originally revealed that transient stability binary classification by deep Lyapunov learning (DLL) is equivalent to constructing a semi-analytical LF in the state space. Case studies validate the effectiveness of the proposed DLL scheme.