Fast Reliability Assessment of Power Systems Based on Interpretable Autoencoder

Abstract

To shift the reliability computational burden from real-time to offline, recent works use a deep learning method for rapid reliability assessment. Most existing data-driven methods are based on training the neural network which is considered a black-box model. In lack of interpretability, it fails to provide convincing information for operating staff. In this regard, this paper proposes an interpretable autoencoder method for reliability assessment. First, a deep network is constructed for rapid calculation of the system reliability, and a weight initialization method is proposed based on feature reconstruction. Thereafter, the model is interpreted based on partial dependence (PD) functions to map relationships between reliability and power injections. Additionally, the Gaussian noise strategy is designed for input features. The proposed method is tested in the RTS-79 system. The partial dependence of reliability can be analyzed with the proposed approach.