Machine Learning Improves Power-System Contingency Security Screening

Summary

The study investigates whether machine-learning classifiers can support faster security screening for power-system contingencies. It assigns contingency scenarios to safe, moderate, or severe classes, using Newton-Raphson load-flow calculations to derive system data and the Overall Performance Index as the safety measure. The researchers address class imbalance with SMOTE and reduce dimensionality with Principal Component Analysis, then compare K-Nearest Neighbours, Random Forest, and Support Vector Machines. Experiments use datasets generated from N-k contingencies with k equal to 1, 2, and 3 on the IEEE-14 and IEEE-30 bus systems. Four preprocessing variants are tested: normalization, SMOTE balancing, PCA transformation, and the combination of SMOTE with PCA. Precision, recall, and F1 score are used for evaluation, with particular emphasis on detecting severe contingencies. Random Forest performs best overall, reaching an F1 score of 0.97 on IEEE-30 and 0.86 on IEEE-14. PCA substantially benefits SVM and contributes more to overall performance than SMOTE in the reported experiments, while SMOTE improves recall but can also increase false positives. KNN performs best with SMOTE and PCA transformations. The authors present machine learning as a potentially scalable alternative to traditional contingency analysis for real-time security assessment, while the results remain tied to the tested IEEE benchmark systems and experimental configurations.