Title:research On Feature Extraction And Adaptive Recognition Methods For Series Arc Fault In The Low-voltage Dc Distribution Network

Under the goal of "carbon neutrality" and "carbon peak" in China,DC green buildings have broad development prospects.However,poor contact,loose joints,and broken insulation of wires could occur in the low-voltage DC distribution system inside buildings,which can cause DC series arc fault and seriously endanger the safe and stable operation of the low-voltage DC distribution network.At the same time,the low-voltage DC distribution system in buildings has the characteristic of strong time-varying in power source and load,making the electrical characteristics of DC arc fault more complex.Therefore,it is urgent to study adaptive recognition methods for DC arc fault,in order to achieve accurate recognition of variable arc fault in DC building scenarios.This thesis firstly introduces the concept and electrical characteristics of DC arc fault,and conducts DC arc experiments in a simulated environment of building DC distribution networks.The obtained low-voltage DC distribution network arc fault signals are analyzed from the changes in current amplitude in the time domain and the wavelet packet energy coefficient of arc fault signals in the time-frequency domain.Based on the comparison of evaluation indicators,the optimized wavelet basis function for time-frequency analysis and the significant feature frequency band for arc fault recognition are determined.A method for extracting time-frequency feature signals of DC series arc fault is studied,and a DC series arc fault recognition model based on convolutional neural network is constructed using time-frequency feature signals as the basis for arc fault recognition.Exploring the small sample training situation,it is found that the pre-trained recognition model may experience recognition failure under various unknown working conditions.Therefore,transfer learning is proposed to adaptively update the parameters of the recognition model of convolution neural network,so as to realize the adaptive recognition of the changeable arc fault.An adaptive arc fault recognition method for DC series fault based on transfer learning is studied.In the face of the difference caused by the system operating conditions in the feature distribution between the unidentifiable working condition and the model pre-training working condition,the similarity measurement method is used to measure the difference,and combined with the probability distribution regularity of the recognition model output,the effective distinction between the identifiable working condition and the unidentifiable working condition is realized,which provides the conditions for transfer learning to update the arc fault recognition model.The Domain Adaptation method with the maximum classifier difference is used to achieve the best model update effect among different transfer learning methods.Then the model update time is further shortened by simplifying the features of the migrated data.In view of the complexity of transfer learning in identifying the updated conditions of the recognition model,an adaptive recognition method of DC series arc fault based on continuous learning is further studied.By combining a gated automatic encoder with a progressive neural network,the best model updating effect among different continuous learning methods is achieved.Combined with task correlation analysis,the arc fault recognition model could effectively face multiple unknown working conditions.Based on the research of series arc fault recognition algorithms,a low-voltage DC distribution network arc fault detection and recognition system is constructed using an embedded computer,Raspberry pie.Performance testing and effectiveness verification are conducted on two series arc fault recognition algorithms.The results show that the DC series arc fault recognition algorithm based on continuous learning has better recognition performance,but its model update time is long and occupies a large storage space.