Analytic Deep Learning and Stepwise Integrated Gradients-Based Power System Transient Stability Preventive Control

Abstract

Trajectory sensitivity (TS) is a powerful alternative to linearize strong-nonlinear transient stability patterns. However, due to its incomplete view of an intact power system operational space, it can be readily to drive preventive control to stuck in local stationary solution. To mitigate this issue, an integrated gradient (IG)-based stepwise control method is proposed. The key idea is to develop IG to enable a control sensitivity taking both merits of local and global sensitivity, further help preventive control escape from local stationary points. Particularly, deep belief network (DBN) is firstly utilized to parameterize pattern between transient stability index (TSI) and steady-state dispatches. It is then leveraged to infer IG to linearize transient stability constraints in real-time. Interior point method is employed to solve the linearized model. A stepwise control strategy is finally proposed to adaptively tune the integrated interval of the applied IG, such that the preventive control path can be ensured to follow the trend of improving transient stability. Case studies on two benchmarks show that the proposed method outperforms traditional TS method in terms of efficiency and economy, and provides a promising way to enhance engineering sensitivity methods.