Transfer Learning for Power Systems Protection Applications

This paper presents a transfer learning approach for protection and relaying applications in power systems. Transfer learning allows for the use of pre-trained deep learning models, enabling them to classify on new problems of the same domain. The strategy deployed here is to use scalogram (wavelet transform) of current signal as an image feature and using a pre-trained image classification model AlexNet to perform fault classification task. The proposed transferred model has superior accuracy compared to a custom built deep-learning classifier and requires minimal dataset and computation resources for training. The technique is tested on all categories of faults and is found promising for further investigation.