Operational Machine Learning Approach for Detection of High Impedance Faults in Distribution Power Networks

Abstract

Abstract Due to insufficient quantitative and indistinguishable qualitative disturbance caused by the High Impedance Faults (HIFs), the detection of the same has been challenging and a topic of active research for decades. The article presents the novel Operational Machine Learning (MLops) concept for HIF detection. The proposed MLops algorithm is combined with the conventional current thresholding-based principles to reduce the ML models’ search space, resulting in improved detection accuracy. Moreover, due to the inherent adaptability of the MLops technique, the proposed method is found to be better capable of generalization as the fault current signature varies. In addition, a novel cumulative sequential ML ensemble method of prediction is used to eliminate the problem of false positive HIF detection without compromising the sensitivity of the HIF detection. The article also describes the method to accommodate for the maximum variance caused by the HIF by using multiple models of HIFs to generate the synthetic data. The synthetic data is obtained using minimal data from the experimental hardware set-up for validation purposes hence improving the utility of the proposed ML-based method in practical scenarios. The ML models such as Support Vector Machines (SVM), Decision Tree (DT), and K-nearest neighbors (KNN) are trained on the synthetic data. The proposed MLops algorithm is tested on the experimental set-up operating at 400V (L-L). The results for the detection accuracy in noisy environments and adaptability due to the parameter update confirm the desired objectives of achieving high sensitivity and robustness for HIF detection.