2022 IEEE International Conference on High Voltage Engineering and Application (ICHVE 2022)

With the diversified development of electricity demand, non-linear electrical equipment continues to increase. Fault arcs have become an important cause of electrical fires. However, the time-frequency domain characteristics of arc faults of different types of loads are not identical; even for the same type of load, there may be differences in the time-frequency domain characteristics of arc current at different times. In practice, it is difficult to enumerate the time-frequency domain characteristics of arc faults under each load condition. The traditional arc fault protection method based on arc characteristics and preset thresholds cannot meet the new requirements of arc fault identification and protection under the condition of variable nonlinear load conditions or unknown load. In this paper, an intelligent identification model of arc fault based on improved AlexNet algorithm is proposed, which realizes the accurate identification of arc fault of unknown load type.

In this paper, the structure of the AlexNet model is optimized, and two 3×3 small convolution kernels are stacked to replace the 5×5 large convolution kernels in the AlexNet model, saving 28% of the parameters. The stochastic gradient descent method is adopted to optimize the weight parameter update process, which reduces the learning difficulty of the model and shortens the training time. At the same time, the training strategy of adaptive learning rate adjustment is adopted to speed up the convergence speed of the network. Using the training set, validation set, test set and data set divided in different proportions to train the network and verify its performance, the optimal hyperparameter combination is selected.

The results show that the improved AlexNet fault arc identification model can directly extract the nonlinear load arc fault characteristics from the current waveform using the deep learning process, and distinguish the normal load condition from the fault arc. This method can realize the accurate identification of the fault arc irrespective of the characteristic quantity and the load type, which is suitable for the identification of the unknown load fault arc. After verification, the improved AlexNet fault arc identification model has an accuracy of 98.5% for known load fault arc identification and 97.5% for unknown load fault identification.

In this paper, an intelligent identification method of arc fault based on improved AlexNet algorithm is proposed. With stronger adaptability and a high identification accuracy, this method does not need to manually determine the threshold value and can realize the identification of the unknown load fault arc.