Comparative assessment of machine learning techniques for fault-induced voltage sag classification

Objective: To compare various artificial intelligence techniques for classifying fault-induced voltage sags using quantitative and qualitative criteria, alongside the analytic hierarchy process. Methodology: Extensive synthetic signals of fault-induced voltage sags were generated using MATLAB/Simulink simulations. Feature engineering included the application of transformations such as the space phasor model, discrete Fourier transform, and short-time Fourier transform. These transformations enabled the extraction of time-series, spectral, and statistical features from the signals, generating ten different feature sets. The most relevant features were then selected using several algorithms to optimize classification performance. Decision trees, support vector machines, and artificial neural networks were employed for classification, with performance evaluated based on computation time, storage requirements, accuracy, and interpretability. The analytic hierarchy process was applied to assess the overall suitability of each approach. Results: Decision trees demonstrated speed, accuracy, and high interpretability, making them ideal for real-time applications. Support vector machines also achieved good accuracy but required more resources and had moderate interpretability. Artificial neural networks offered balanced performance with limited interpretability. Conclusions: Among the evaluated algorithms, decision trees are the most suitable for real-time classification of fault-induced voltage sags. However, the choice of technique should align with specific application needs. Future research should consider additional criteria and focus on improving interpretability through explainable artificial intelligence techniques.