Identifying critical components in power networks through machine learning

As modern power systems become more complex, identifying critical components for their reliable operation is becoming increasingly important, especially in dealing with widespread cascading outages from large-scale disturbances. In this context, this paper introduces a novel method for identifying critical components in terms of their contribution to cascading-induced demand not served (DNS), utilizing feature selection techniques integrated with a machine learning classifier model. While existing methods predominantly identify critical assets using DNS, they often overlook the effects of cascading failures. Furthermore, those that do consider cascading failures typically focus solely on line overloading. This work, therefore, introduces a machine learning-based methodology coupled with a cascading failure model, considering protection mechanisms like excitation limits, under/over frequency, and line overloading. Tested on the IEEE 24-bus system, the validation method exhibits strong sensitivity (80.35%-90.03%) and specificity (80.35%-94.10%). Additionally, SHapley Additive exPlanations (SHAP) is used for validation by comparing its results with the proposed method.