DC arc faults in photovoltaic (PV) systems often create serious hazards, so that they need to be detected quickly and accurately. Compared with the traditional time and frequency domain arc fault detection methods, the method combined with machine learning achieves higher detection accuracy under different working conditions. In view of the above-mentioned background, this paper designs an arc fault online detector that can be integrated into PV systems, based on neural network (NN) algorithm. First, an arc fault detection experiment platform is built. By analyzing the measured arc data, the characteristic frequency band of the arc signal is selected. On this basis, a filter circuit is designed. After the filtered signal processed by digital signal processing (DSP), it is transmitted through the serial port to the host computer to complete the data acquisition. Then, the method of selecting, calculating and preprocessing the features of the signal is proposed. Using the processed features, the NN classifier is trained and optimized in the host computer to achieve a classification accuracy of over 99.5%. Finally, the trained parameters are used in the detection code of DSP, and the arc detection can be completed within 4ms, realizing the function of real-time arc fault detection.

photovoltaic (PV) system;dc arc;online fault detection;feature extraction;neural network (NN)