Abstract
This paper presents an exhaustive machine learning model for fault resistance estimation and classification in 11 kV distribution networks. In MATLAB/Simulink, a high-fidelity Simscape Power Systems model is built around an 11 kV, 30 MVA source connected to two 100 km PI-section transmission lines, a midpoint-fault bus, and a downstream 1 MVA, 11kV/0.4 kV transformer-Monte Carlo simulation. An automated Monte Carlo simulation is used to generate a dataset of 5000 Phase-A single-line-to-ground fault scenarios, with fault resistances uniformly distributed from 0.01 to 50 Omega. Three-phase voltage and current waveforms are extracted and converted into 20 physical, interpretable time-domain features. Three machine learning models - Gradient Boosting (GB), Random Forest (RF), and Support Vector Machine (SVM) - are trained and benchmarked on both regression and classification tasks. The Gradient Boosting regressor achieves a root mean square error (RMSE) of 0.0276 Omega and an R2 of 1.0000 on a held-out test set of 1000 samples. In contrast, the Random Forest classifier achieves 99.90 per cent accuracy for fault severity classification. The noise-robustness analysis at signal-to-noise ratios (SNRs) between 10 and 50 dB shows that RF is more favourable for use in noisy environments, with an RMSE of 0.96 Omega and 96.72 per cent accuracy at SNR = 50 dB. A multi-fault-type extension to seven fault configurations (AG, BG, CG, ABG, BCG, AB, ABCG) over 3500 simulations demonstrates 100% fault-type identification accuracy and pertype regression R-squared exceeding 0.996 for all fault types. These results establish a robust, simulation-driven machine-learning pipeline suitable for deployment in adaptive protection and fault-management systems.