Researchers have published a machine learning method for contingency screening in power systems, achieving an F1 score of up to 0.97 on the standard IEEE-30 test system. The study, published on arXiv (identifier 2609.04300), proposes a pipeline combining data-optimization techniques to support proactive power system decision-making and mitigate large-scale grid failures.
The research method classifies the security status of power systems during contingencies into three classes: safe, moderate, and severe. To generate the dataset, the authors applied the Newton-Raphson load flow method to extract operational parameters from N-k contingency scenarios with k equal to 1, 2, and 3. The Overall Performance Index (OPI) was used as the safety metric across two standard test networks: the IEEE-14 and IEEE-30 bus systems.
For data preparation, the study evaluated four hybrid preprocessing configurations: normalized data, SMOTE-balanced data (Synthetic Minority Over-Sampling Technique), PCA-transformed data (Principal Component Analysis), and combined SMOTE and PCA transformation. Three machine learning models—Random Forest (RF), Support Vector Machines (SVM), and K-Nearest Neighbours (KNN)—were trained and evaluated on precision, recall, and F1 score, with priority placed on severe contingency classes.
According to the published paper, Random Forest achieved the highest performance with an F1 score of 0.97 on the IEEE-30 system and 0.86 on IEEE-14. SVM benefited significantly from PCA dimensionality reduction to enhance classification accuracy, while KNN performed best under the combined SMOTE and PCA configuration. Notably, findings indicate that PCA contributes more to overall model performance than SMOTE. While SMOTE improves recall for severe contingencies, it also introduces false positives, creating a trade-off with general accuracy.
The study highlights machine learning as a viable and scalable alternative to traditional contingency analysis for real-time power system security assessment. However, the authors have not yet provided validation on real-world utility-scale power grids beyond the IEEE-30 testbed, nor computational latency metrics under live dispatch conditions.