Research On Transient Stability Assessment Method Based On Deep Learning And Its Prediction Uncertaint

Transient stability assessment(TSA)is a critical task for ensuring the secure operation of power systems.It aims to determine whether the power system can return to a new steady state after a disturbance.With the increasing scale and complexity of modern power systems,TSA faces new challenges due to the high-dimensional nonlinearities and uncertainties involved in the system dynamics.Traditional methods based on time-domain simulation and direct methods are often computationally intensive and require accurate system models,which may not be suitable for online evaluation.Therefore,there is an urgent need for a fast,accurate and reliable method for TSA to assist operators in making decisions and improving the operational stability of power systems.Deep learning techniques have shown great potential for TSA due to their ability to learn complex patterns from data without relying on physical models.Moreover,with the wide deployment of wide area measurement systems(WAMS)in power systems,large amounts of power data can be collected and stored,providing data support for the application of deep learning in TSA.Based on deep learning theory and uncertainty quantification methods,this paper proposes a transient stability assessment framework based on convolutional neural networks and their variants,which can achieve end-toend transient stability assessment and uncertainty quantification of prediction results.The main contributions and novelties of this paper are as follows:This study proposes a transient stability evaluation method based on onedimensional convolutional neural network(1D-CNN)that uses the underlying measurement data of PMU as the input features without manual feature selection.The1D-CNN model can capture the temporal information of the input features and achieve end-to-end temporal feature extraction and transient stability assessment.Moreover,to address the issue of prediction accuracy degradation caused by sample imbalance in the dataset,this study applies the SVM-SMOTE technique to balance the samples.The simulation results show that the proposed method can achieve online assessment accuracy of 99.31%,rapidity of 0.27 s,and reduce the underestimation of unstable samples in unbalanced data sets.This study constructs a transient stability assessment model based on Bayesian convolutional neural network(BCNN),which introduces uncertainty estimation by placing a distribution over the parameters of the CNN network.The BCNN model can predict the transient stability state of the power system based on the underlying PMU measurement data and provide uncertainty estimates of the prediction results,which enhances the interpretability and reliability of the BCNN model.Operators can set different uncertainty thresholds to improve the prediction results according to their specific needs,making the prediction results more instructive.The experimental results show that the model is more robust to noisy data and feature-deficient data,has good adaptability to small data sets,and the uncertainty features of the output can well reflect the confidence level of the model in the predicted samples.