Power System Transient Stability Prediction with a CNN Trained with Steady-State Input Variables

Using steady-state information, a Convolutional Neural Network (CNN) model is presented in this paper to assess the transient stability of a test power system according to Stable and Unstable categories. A set of representative input/output examples were generated under a wide range of operating conditions with the system subjected to a three-phase fault at every involved transmission line. Input observations such as predisturbance values of active and reactive power of transmission lines, and voltage magnitudes at the terminal nodes of these lines, were collected and then associated with target classification labels. Corresponding categories were determined from time domain simulations and the calculation of a transient stability index based on the resultant post-disturbance behavior of machine power angles. With the gathered information, a CNN was trained and evaluated employing a stratified k-fold cross-validation process. To boost the performance of the proposed CNN, an approach where the trainable parameters of front layers are frozen during the learning phase was implemented and explored. Simulation results in terms of common classification metrics show the enhanced performance of the proposed model with the layer freezing strategy, as compared to the base CNN considered in the studies.