Modeling failures in smart grids by a bilinear logistic regression approach.

Area of Science:

• Electrical Engineering
• Computer Science
• Data Science

Background:

• Modeling complex systems for event recognition is challenging, especially when balancing accuracy with explainability.
• Existing machine learning models often lack interpretability, hindering their application in critical infrastructure like power grids.

Purpose of the Study:

• To develop an explainable and accurate machine learning model for power grid failure analysis.
• To correlate external events with power grid characteristics for better understanding of fault phenomena.
• To enable predictive maintenance, condition monitoring, and risk assessment through an interpretable AI paradigm.

Main Methods:

• Utilized a data-driven approach with a bilinear logistic regression model.
• Grounded the model in a specific neural architecture, creating a white-box system.
• Trained the model on a real-world dataset of power grid failure data.

Main Results:

• The bilinear white-box model achieved performance comparable to existing classifiers in identifying faulty states.
• The model's low computational complexity facilitated insights into fault phenomena and event correlations.
• A vulnerability vector was estimated for power grid components, serving as an interpretable 'label'.

Conclusions:

• The proposed bilinear model offers a powerful tool for accurate and explainable power grid failure analysis.
• It effectively uncovers relational information between exogenous causes and grid characteristics.
• The model supports advanced applications like predictive maintenance, risk assessment, and scenario analysis within the explainable AI framework.