This paper introduces a deep learning approach for addressing fault detection and location issues in power distribution grids. The proposed method utilizes a 20-layer deep neural network (CNN) that employs the ReLU activation function in the hidden layers and the Softmax function in the pre-classification layer. The network’s input layer dimensions are 224 x 224 pixels, with each input image being a composite of seven images, each sized 32 x 224 pixels. Data is gathered using continuous wavelet transform and Hilbert transform on each signal, followed by feature extraction and conversion into RGB images. The signals include three-phase voltages, currents, and zero-sequence voltages, collected across different locations, times, and fault resistances for use in the training and testing of the deep neural network. The output layer categorizes faults into ten classes, covering types such as single-phase to ground, two-phase, and three-phase faults. Additionally, the location identification layer categorizes faults into five classes corresponding to five zones on a specific feeder. The proposed method is validated using a test power system simulated in MATLAB 2020b, which includes six main feeders and two sub-feeders. The validation is conducted under two scenarios: one without distributed generation sources and with data noise at SNR = 10dB, and another with distributed generation sources and SNR = 40dB. The method demonstrated strong performance in all cases, and the trained network successfully detected and located faults in previously untested areas.
Read full abstract