A robust deep learning framework for smart grid stability classification using statistical validation

The classification of smart-grid stability is difficulty due to nonlinear interactions among the participant reaction-time, elasticity of power consumption and electricity-price. This paper proposed a comparative deep-learning model of classifying stable and unstable operating conditions based on 10,000 observations of simulated four-node electrical-grid stability dataset. To avoid deterministic redundancy and target leakage, we omitted p1 and the continuous stability variable stab, making 11 input predictors. They were data-quality assessment, training, standardization, stratified data partitioning and five-fold cross-validation. Four traditional machine-learning models (Logistic Regression, RBF-SVM, Random Forest and Gradient Boosting) and five deep-learning models (DNN, AutoEncoder-DNN, 1D-CNN, CNN1D-LSTM and FT-Transformer) were tested using a similar conditions. The DNN achieved the best overall test performance, with an accuracy of 97.10%, precision of 97.28%, recall of 98.20%, F1-score of 97.74%, ROC-AUC of 0.9963 and PR-AUC of 0.9980. AutoEncoder-DNN and RBF-SVM had the accuracy of 96.90% and 96.05, respectively. Precise paired McNemar tests with Holm correction indicated that the DNN performed better in comparison to the other tested networks with the exception of the AutoEncoder-DNN where the difference was not statistically relevant. The analysis of SHAP and feature-group ablation showed that the most impactful predictor groups were reaction times and price-elasticity coefficients of participants. The DNN achieved 95.60% accuracy with standardized Gaussian noise with a value of 0.10, but dropped to 90.90% in the case of 10% of measurements random and median-imputed. These results indicate that the DNN is effective on the studied simulated benchmark. However, it needs to be validated with actual PMU or SCADA measurements and larger and more heterogeneous grid topologies prior to operational implementation.