Year: 2026
Type: article
Abstract: Artificial-intelligence-based smart-grid stability predictors are commonly assessed on clean data, leaving their tolerance to measurement degradation insufficiently characterized. This study compares Logistic Regression, Random Forest, XGBoost, SVM-RBF, and an optimized deep neural network using the Electrical Grid Stability Simulated Data. The data were stratified into 70% training, 15% validation, and 15% testing subsets. Trained models were evaluated on a fixed test set under Gaussian noise, Bernoulli–Gaussian impulsive noise, and measurement dropout at five severity levels and 30 paired realizations without retraining. Robustness was quantified using the Matthews correlation coefficient, ROC-AUC, false-negative rate, area under the robustness curve, and paired Friedman and Wilcoxon–Holm tests. SVM-RBF achieved the strongest clean performance and the highest MCC AURC under Gaussian noise and measurement dropout, whereas XGBoost achieved both the highest MCC AURC and the lowest FNR AURC under impulsive noise. Under equal disturbance-family weights, XGBoost obtained the highest mean MCC AURC of 0.843542 and the lowest mean FNR AURC of 0.043506, and it achieved the highest weighted MCC AURC in 78.66% of the evaluated weight combinations. Model selection should therefore reflect both the expected measurement disturbances and the operational metric priorities rather than clean accuracy alone.
Institution: Universitas Muhammadiyah Malang
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $13
OpenAlex ID: W7225171268