A DEEP LEARNING MODEL FEATURE INTERPRETABILITY ANALYSIS METHOD FOR POWER SYSTEM TRANSIENT STABILITY ASSESSMENT

The mechanism underlying power system transient stability is complex. Deep learning models offer an effective solution for capturing complex mapping relationships, making them widely employed in transient stability assessment. However, the deep learning models face challenges in ensuring the effecti...

Full description

Saved in:
Bibliographic Details
Published inScientific Bulletin. Series C, Electrical Engineering and Computer Science no. 4; p. 305
Main Authors Zho, Yibo, Zhang, Liang
Format Journal Article
LanguageEnglish
Published Bucharest University Polytechnica of Bucharest 01.01.2023
Subjects
Accuracy
Deep learning
Feature extraction
Knowledge acquisition
Machine learning
Security
Stability analysis
Transient stability
Online AccessGet full text

Cover

More Information
Summary:The mechanism underlying power system transient stability is complex. Deep learning models offer an effective solution for capturing complex mapping relationships, making them widely employed in transient stability assessment. However, the deep learning models face challenges in ensuring the effectiveness of feature extraction due to the lack of domain knowledge support. This limitation hampers improvements in evaluation accuracy. Furthermore, the inability to comprehend the acquired knowledge of the model raises concerns about trusting evaluation results, especially in security-sensitive scenarios. To address these issues, this article proposes a method for analyzing the interpretability of deep learning model features in power system transient stability assessment. Firstly, we construct a CNN model specifically designed for transient stability assessment. Then, we introduce a global interpretation method known as maximizing activation (AM) to obtain a comprehensive interpretation of typical stable modes within the model's injection space. Finally, the Class Activation Map (Grad-CAM) is utilized to identify dominant features in the injection space, providing guidance for the online application of transient stability assessment. The case studies show that this method can make operators easily understand the transient stability assessment knowledge learned by neural networks and improve the accuracy of transient stability assessment under the security region.
ISSN:2286-3540