Fault Classification and Localization in Series-Compensated Transmission Lines Using Wavelet Features and Machine Learning

This study evaluates the effectiveness of wavelet-derived features for fault classification and fault location estimation in a three-phase series-compensated 735-kV, 300-km transmission line. Eleven fault types were simulated under varying locations and inception times, and three-phase voltages and currents were sampled at 1 MHz. The dataset consists of a total of 1000 simulation cases. The case results guarantee sufficient diversity for feature evaluation to cover random variations in fault resistance, location, and inception angle. From each phase and signal type, a set of CWT features—including peak time, peak frequency, peak energy, total energy, mean coefficient magnitude, coefficient standard deviation, and wavelet entropy—was extracted. These features were subsequently employed as inputs to six classification algorithms (Random Forest, Logistic Regression, SVM-RBF, KNN, Gaussian Naive Bayes, and MLP) and six regression models (Linear, Decision Tree, Random Forest, SVR-RBF, KNN, and MLP). Model evaluation was performed using an 80–20 train–test split and 5-fold cross-validation to ensure statistical robustness. Results show that wavelet-based features provide strong discriminatory power for multi-class fault analysis. The MLP and Logistic Regression models achieved 93.00% and 92.50% accuracy on the training set, while maintaining generalization under cross-validation. In regression, the Random Forest and MLP Regressor models provided the most accurate fault location estimation, achieving R² scores of 0.953 and 0.945, respectively. Discussion of model robustness, parameter justification, and practical implementation feasibility is also included.