Abstract
Accurate estimation of fault resistance and reliable discrimination of fault type are foundational to the protection and self‐healing capability of medium‐voltage distribution networks, yet both tasks are severely confounded by high‐impedance faults (HIFs) whose low and variable current signatures evade conventional overcurrent and distance protection. This paper presents a noise‐resilient machine learning framework that performs continuous fault resistance regression and multiclass fault‐type classification on a simulated 11‐kV radial distribution feeder. A high‐fidelity electromagnetic model was constructed in MATLAB/Simulink–Simscape, comprising a 35 MVA Thévenin‐equivalent source, two cascaded PI‐section lines totaling 10 km, a 5‐MVA distribution transformer, and an aggregated downstream load. An automated Monte Carlo pipeline generated 6000 single‐phase‐to‐ground (AG) fault cases, with fault resistance log uniformly sampled over 0.01–500 Ω, and a further 3500 cases spanning seven canonical fault types. From pre‐ and postfault voltage and current windows sampled at 10 kHz, 22 physically interpretable features were extracted, encompassing faulted‐ and healthy‐phase magnitudes, asymmetry indices, zero‐sequence components, and derived impedance quantities. Gradient boosting machine (GBM), random forest (RF), and support vector machine/regression (SVM/SVR) models were trained and benchmarked under clean and noise‐contaminated measurement conditions with additive white Gaussian noise spanning 10–50 dB SNR. Under clean conditions, the GBM regressor achieved
R
2
= 0.9961, RMSE = 8.73 Ω, and MAE = 4.21 Ω; the RF classifier attained 99.34% accuracy on fault‐severity classification. The framework maintained
R
2
> 0.97 and classification accuracy > 96% down to 30 dB SNR, which serves as a qualitative reference level representative of metering‐class instrument transformer measurement quality, establishing practical sensor requirements for field deployment. Multifault‐type extension yielded 98.91% identification accuracy with per‐type regression
R
2
consistently exceeding 0.98. All results are based on synthetic simulation data from a single feeder configuration, one fault location, and AWGN‐based noise modeling; field validation remains future work. The results demonstrate that interpretable, ensemble‐based learning on physically grounded features can deliver simultaneous resistance estimation and fault typing suitable for integration into intelligent protection and smart‐grid monitoring systems.