Machine learning tools for active distribution grid fault diagnosis

Faults in power distribution networks cause customer minute and economic losses. A crucial part of the protection system of such grids is effective fault diagnosis for the acceleration of the restoration process after being subjected to any faults. This article presents the fault diagnosis approach for the active distribution grid, and the method consists of machine learning tools and signal processing techniques. The Hilbert-Huang transform (HHT) and discrete wavelet transform (DWT) are considered as the signal processing tools whereas the feedforward neural networks (FFNN) are considered as the machine learning tools. The extracted features using the signal processing tools are fetched into the neural network models for the development of fault detection, fault classification, and either fault location or faulty section identification models. The proposed approach is tested on two different distribution feeders. The first one is a simplified four-node test feeder modeled in MATLAB/ SIMULINK environment. In contrast, the second one is the IEEE 13-node distribution network with the incorporation of three renewable energy resources (RER) modeled in the Real-Time Digital Simulator (RTDS) machine. The uncertainty, e.g., RER generation, load demand, and fault information, associated with the test feeders are modeled using different probability density functions. Obtained results demonstrate the efficacy of the proposed models for both noise-free and noisy data. Finally, the developed models show their independence in the variation of the pre-fault loading conditions, fault inception angle, and fault resistance.

Automated summary

This article presents a fault diagnosis approach for active distribution grids using signal processing techniques and machine learning tools. The approach combines the Hilbert-Huang transform and discrete wavelet transform as signal processing tools, while feedforward neural networks are used as machine learning tools. The proposed approach is tested on two different distribution feeders and shows effectiveness in detecting, classifying, and locating faults. The developed models also demonstrate independence in varying pre-fault conditions, fault angles, and fault resistance.