Anomaly Detection Using LSTM-Based Variational Autoencoder in Unsupervised Data in Power Grid

This article proposes a deep generative model for anomaly detection in unsupervised power grid data. One-class classifier-based methods suffer from performance degradation when training data contain anomalous samples. Due to the temporal characteristics in most of the power grid datasets, we explore a long short-term memory-variational autoencoder-based deep generative model that can tolerate the moderate presence of anomalous data during training instead of standard data. This work demonstrates the advantage of reconstruction-based methods over clustering-based methods. As part of the reparameterization of the latent layer, a method is proposed by employing wavelet decomposition of the wavelet coefficients found from the high and medium frequency representations of the input time-series data. For further improvement, we have incorporated a cosh -based cost function instead of the traditional consideration of the L₂ norm-based cost function. The numerical results demonstrate an improvement of performance metrics, such as AUC by 0. 1–0. 2 of our method over other benchmark methods. The transient stability threshold () is an important system parameter in the performance assessment of power grid systems. Through time domain simulations, it has been shown that =0. 3 obtains optimal accuracy for transient stability assessment in the IEEE NewEngland-39 bus.