Siamese recurrent neural networks for the robust classification of grid disturbances in transmission power systems considering unknown events

Abstract

The automated identification and localisation of grid disturbances is a major research area and key technology for the monitoring and control of future power systems. Current recognition systems rely on sufficient training data and are very error-prone to disturbance events, which are unseen during training. This study introduces a robust Siamese recurrent neural network using attention-based embedding functions to simultaneously identify and locate disturbances from synchrophasor data. Additionally, a novel doublesigmoid classifier is introduced for reliable differentiation between known and unknown disturbance types and locations. Different models are evaluated within an open-set classification problem for a generic power transmission system considering different unknown disturbance events. A detailed analysis of the results is provided and classification results are compared with a state-of-the-art open-set classifier.

Siamese Sigmoid Networks for the open classification of grid disturbances in power transmission systems

“…Based on the preliminary work in ref. [37], the SSN consists of two model branches with shared parameters to estimate the class affiliation of a query instance X Q by the comparison with multiple support instances X S with a known class affiliation y . The basic model architecture is given in Figure 4.…”

“…In contrast to ref. [37], the fully trainable SSN approach allows to automatically learn class‐dependent threshold distances and margins m , which leads to compact decision boundaries around the classes with a reduced open‐set risk—see Figure 2.…”

“…In particular, the paper’s contributions can be summarised as follows: extends the double‐sigmoid classifier approach from ref. [37] to automatically learn class‐dependent threshold distances and margins with an additional selection vector and a custom penalty loss formulation allows the consideration of varying intra‐class and inter‐class variances in large classification problems introduces an efficient multiple‐negative sampling training strategy to improve the model convergence and accuracies evaluate SSNs against multiple benchmark methods for different closed‐set and open‐set disturbance classification tasks …”

“…extends the double‐sigmoid classifier approach from ref. [37] to automatically learn class‐dependent threshold distances and margins with an additional selection vector and a custom penalty loss formulation…”