Performance analysis of a DNN classifier for power system events using an interpretability method

Autor: Meng Wang, I.C. Decker, Daniel Dotta, Joe H. Chow, Orlem L. D. Santos
Rok vydání: 2022
Předmět:
Zdroj: International Journal of Electrical Power & Energy Systems. 136:107594
ISSN: 0142-0615
DOI: 10.1016/j.ijepes.2021.107594
Popis: Nowadays, there is a clear need for Machine Learning methods capable of extracting relevant and reliable information from synchrophasor data. In this paper, the application of an explainable data-driven method is carried out in order to inspect the performance of DNN classifier for event identification using synchrophasor measurements. The DNN classifier is the Long–Short Term Memory (LSTM) which is suitable for the extraction of dynamic features. The key advantage of this approach is the use of an interpretability inspection named SHAP (SHapley Additive exPlanation) values, based on cooperative game theory (Shapley values), which provide the means to evaluate the predictions of the LSTM and detecting possible bias. The main contributions are stated as follows: (i) it explains how the LSTM classifier is making its decisions; (ii) it helps the designer to improve the training of the classifier; (iii) certify that the resulting classifier has a consistent and coherent performance according to domain knowledge of the problem; (iv) when the user understands that the classifier is making coherent decisions, it clearly reduces the concerns of the application of DNN methods in critical infrastructure. Additionally, the proposed approach is evaluated using real synchrophasor event records from the Brazilian Interconnected Power System (BIPS).
Databáze: OpenAIRE