The detection of single phase‐earth faults has been a difficult task for a long time due to its very low current in high impedance grounded fault especially in a neutral un‐effectively grounded system. To address this issue, this paper firstly proposes multi‐learner based single phase‐earth fault identification models. First, to erase disturb noise in fault recording profiles, a denoising model is put forward based on the wavelet transform optimized via the proposed threshold improvement approach. Second, feature engineering reflecting local and/or global evolutionary process of fault evolving features is modelled, and in turn two key feature transforming techniques of principal component analysis (PCA) and random forest (RF) are individually employed to recognize the best effective combination of fault feature set. Subsequently, six learners of logistical regression (LR), support vector machine (SVM), K‐neighbor (KN), RF, XGBoost and LightGBM based fault identification models are individually custom‐designed which the feature subset in high priority are fed into. Furthermore, to guide the model optimization, several advanced manners of hyper‐parametric sampling via normal and Chi‐2 distribution, learning curve, validation curve, receiver operation characteristic curve (ROC) are applied. Numerical studies indicates the the high value of the proposed model when applied in engineering practice.

Single-phase earth ground faults are the most frequent faults likely to occur but hard to identify in a distribution system, especially in a neutral ineffectively grounded system. Targeting on this goal, a novel AdaBoost-based single-phase earth ground fault identification model is put forward. First, after depicting the zero-sequence circuit of the distribution system, a feature engineering that can reflect local and global evolutionary processes in the fault period is constructed in detail. Second, to overcome two problems, namely, different number problems between fault and non-fault samples and curse of dimension, principal component analysis is used for feature extraction, in which only a small number of low-dimension mapped features are extracted, and then transmitted into the AdaBoost-based ground fault identification model. Subsequently, this work borrows from machine learning and applies its learning curve and receiver operating characteristic curve to guide the optimization of the proposed identification model. Numerical studies verify the effectiveness and adaptability of the proposed model toward solving single-phase earth ground faults.

No abstract