Comparison of advanced signal decomposition techniques for the classification of PQDs by machine learning algorithms

Accurate detection and classification of Power Quality Disturbances (PQDs) are critical for ensuring reliability in modern distribution networks with high renewable energy penetration. This study evaluates four signal decomposition methods—Empirical Mode Decomposition (EMD), Ensemble EMD (EEMD), Complete Ensemble EMD with Adaptive Noise (CEEMDAN), and Variational Mode Decomposition (VMD)—in conjunction with a Random Forest Classifier (RFC) for PQD classification. The IEEE-1159 synthetic benchmark dataset comprising fifteen single and multiple PQD classes was used for training and testing, with 5-fold cross-validation employed to ensure robustness. Hyper parameter tuning of the RF model was performed using grid search to optimize the number of trees, depth, and feature selection strategy. Among the methods, VMD + RFC consistently outperformed the EMD-family techniques, achieving a confusion matrix–based accuracy of 99.16% and a cross-validation accuracy of 94.6% ± 1.42 (95% CI). Paired t-tests confirmed that the accuracy improvements of VMD over the other decomposition methods were statistically significant (p < 0.05). Beyond synthetic benchmarks, the proposed study was validated on a field dataset of the university campus in India collected at the point of common coupling (PCC) where a 500 kWp photovoltaic (PV) system is integrated. Here, VMD + RFC predictions are closely aligned with the reference PQA logs, demonstrating strong generalization capability. The results establish VMD + RFC as a robust and computationally efficient study for PQD classification, combining superior accuracy, statistical significance, and real-world applicability. This contributes both methodological rigor and practical validation, distinguishing the work from prior studies limited to smaller PQD sets or lacking external verification.