Ensemble Learning Model of Power System Transient Stability Assessment Based on Bayesian Model Averaging Method

Abstract

With the increase in the penetration of renewable energy and power electronic devices, the modern power systems will face great challenges for the traditional model-based transient stability assessment (TSA). In this study, a new ensemble learning TSA model based on Bayesian model averaging method is proposed. Firstly, the Long-short Term Memory (LSTM) network and multidimensional deep neural network (DNN) is used to establish eight independent machine learning models, according to the generator power angle, rotor speed electromagnetic power and bus voltage features. Since these sub-models can only achieve different accuracies ranging from 50%, 84.28% to 95% even for the simple test system, it can not have sufficient generalization ability for real world TSA applications. Then the pre-trained sub-models are weighted and averaged by Bayesian model averaging method to obtain the ensemble learning model. Case studies of electromechanical transient simulation are analyzed on the modified IEEE 39-bus test system to verify the effectiveness of the proposed model. The prediction accuracy of the ensemble learning model can reach a stable performance of 94.83%, which has great potential for future online TSA applications.