Abstract: With the rapid development of artificial intelligence algorithms，convolutional neural networks with

fewer layers have been used in distribution network overvoltage recognition. The deep-level network has a higher

recognition rate，but requires a large number of data samples. At present，the amount of data in the existing data set is insufficient to meet the needs of deep-level network training. To this end，a method for establishing distribution network overvoltage data sets required for deep-level network training was proposed. Firstly，the electromagnetic transient simulation software EMTPworks was used to simulate 5 typical overvoltages of 10 kV distribution network and the corresponding JavaScript script was edited，and 16 272 pieces of data were generated by changing the parameters of the fault initial phase angle，transition resistance，line length and other parameters. Then the threephase overvoltage one-dimensional signal was subjected to continuous wavelet transform to obtain a twodimensional time-frequency diagram of the corresponding overvoltage. Afterwards，the two-dimensional timefrequency map was automatically marked according to the characteristics of the original signal，thereby a complete distribution network overvoltage data set was established. Finally，the convolutional neural network（CNN）was used to verify the validity of the 5 types of overvoltage signal data. The results show that the constructed data set has large scale and high validity，and can meet the needs of deep-level network.

      Key words: distribution network；overvoltage；continuous wavelet transform；JavaScript script；convolutional

neural network（CNN）