Research On Fault Line Selection Based On Machine Learning And Wavelet Packet Transform

Fault line selection of distribution network is a research hotspot in recent years.It is of great significance to select the fault line quickly and accurately for the safe and reliable operation of the power system.In view of the shortcomings in the existing fault line selection methods,this paper introduced machine learning,transformed fault line selection into multi-classification problems,and used the classifiers in machine learning to select fault lines.The research work is as follows:(l)Used the four-outline 10 kV distribution network as an example to establish a model,based on the analysis of the steady-state and transient electrical characteristics of each line when single-phase ground fault occurs,determined to use transient current for fault line selection,through a large number of simulation created a dataset for machine learning.(2)Based on the wavelet theory,the wavelet packet transform was used to extract the zero-sequence current of each line,and the energy and modulus maxima of the zero?sequence current in the characteristic frequency band were obtained to select the fault line.The selection results of a large number of examples showed that there is a problem that the mis-selection rate is high because of the criterion that the zero-sequence current of the fault line has the largest energy,and the largest modulus maxima,of which the polarity is opposite when compared to sound lines.A fault line selection method combining machine learning and wavelet packet transform was proposed.The energy and modulus maxima obtained by the wavelet packet transform were used as the dataset used in machine learning,and the random forest was selected as the classifier to construct the fault line selection method.After testing,this method not only inherits the advantages of the method based on wavelet transoform,but also plays the role of random forest classifier,which can reduce the mis-selection rate when the modulus maxima feature is not obvious.(3)A method for fault line selection using a random forest classifier directly with the original zero-sequence current dataset was proposed.Compared with the dataset based on wavelet packet transform,the original zero-sequence current dataset has a higher sample size and an increased amount of information.In order to select the optimal classifier,the training set was used to train the support vector machine,feedforward neural network,decision tree and random forest.The classifying ability was evaluated on the test set,and the random forest shows better performance.After testing,the fault line selection method constructed by random forest has short cross-validation time and accurate line selection,and is not affected by grounding resistance,the distance from the fault point to the busbar,and the initial angular position of the fault.(4)At the end of the article,a system consisting of“offline training”and“online recognition”was designed to provide an application program for machine learning in the field of fault line selection.