A Two-Stage Machine Learning Framework for Predicting Spatial Load Shedding in Cascading Power System Failures

The inherent complexity of power systems makes the precise prediction of cascading failures challenging, compromising their ability to deal with such disastrous events. During stressed conditions (e.g., high demand or high renewable penetration periods), even minor contingencies can trigger widespread cascades resulting in power outages. Accurately forecasting such cascade propagation and resultant load shedding (LS) remains difficult, even with the growing availability of data. Cascades usually propagate through the network by activating complex and interdependent protection mechanisms, where seemingly shallow cascades can lead to significant LS due to the failure of critical components. The same failures under different system conditions, including varying demand levels, can lead to very different cascading outcomes which further enhances the complexity of predicting such events accurately. This study introduces a novel and practical two-stage machine learning-based model for assessing the level of LS with high spatial granularity under varying load conditions. The first stage employs a stacked regression approach to predict spatial LS caused by cascade propagation by modelling nodal LS as interdependent among the buses. The second stage refines these predictions by using a meta-classifier and categorizes the severity of the cascade based on the resulting LS, reducing the errors of the regression model. Together, these stages offer a practical method for improving LS assessment and hence enhancing the timely decision-making against such cascading events. The proposed model is tested on the IEEE 39-bus test network, demonstrating robust performance with a mean F1 score of 73.8%, recall of 71.0%, and precision of 78.2%.