An effective feature extraction method in pattern recognition based high impedance fault detection

High impedance fault (HIF) is problematic in various distribution systems, specially in rural distribution feeders. The fault current of HIF is with low magnitude, non-linear, asymmetrical and random, therefore extracting useful detection features from HIF current and voltage is the key to solve this issue. This paper experiments with 246 conventional electrical features and their combinations and proposes an effective feature set (EFS) via a feature ranking algorithm utilizing simple signal processing technique of discrete Fourier transform and Kalman filter estimation. This EFS is tested in six types of distribution systems and exhibits a promising detection performance in terms of accuracy, dependability and security once a proper pattern recognition classifier is determined. Besides conventional batch learning algorithms, the proposed detection method demonstrates a significant performance in online machine learning environment. Therefore it shows the potential of processing instantaneous signals and updating its prediction model adaptively to detect more HIFs in future smart grid.