Prediction of Cascading Failures in Spatial Networks

Area of Science:

• Computer Science
• Network Engineering
• Reliability Engineering

Background:

• Cascading overload failures pose significant risks to the reliability of large-scale parallel systems.
• Predicting these failures is vital for effective network control and management.
• Analyzing individual node behavior in complex, growing networks is often impractical.

Purpose of the Study:

• To predict the failure rates of nodes within specific groups in a network.
• To leverage spatial-temporal correlations of overload failures for group-based prediction.
• To evaluate machine learning models for their accuracy in predicting cascading overload failures.

Main Methods:

• Simulated overload failure propagation in a weighted lattice network initiated by a central attack.
• Grouped nodes based on shared spatial and temporal characteristics.
• Employed machine learning models: Feedforward Neural Network (FNN), Recurrent Neural Network (RNN), and Support Vector Regression (SVR).

Main Results:

• All tested machine learning models (FNN, RNN, SVR) demonstrated high accuracy in predicting node group failure percentages.
• The models successfully captured the similar failure behaviors of nodes within defined groups during cascading overload propagation.
• Spatial-temporal correlations were effectively utilized for group-based failure prediction.

Conclusions:

• Machine learning offers a viable approach for predicting cascading overload failures in large, complex networks.
• Group-based analysis, informed by spatial-temporal correlations, enhances prediction accuracy.
• The study validates the utility of FNN, RNN, and SVR for network reliability management.