Compression of the Training Sample of the Smart Protection Device without Compromising its Information Capacity

Abstract

In the algorithms of their functioning, Smart Protection Devices use machine learning methods that provide them with the ability to make decisions in the multidimensional space of parameters controlled by the device. To make the Smart Protection Device capable of classifying monitored and alternate modes of the electrical network, it must be trained. Learning is carried out at the development stage, and its effectiveness depends on the training sample. Correctly chosen dimensionality of the precedent space and the power of the training sample guarantee the unambiguity of the solution to the problem of classifying the modes of the electrical system since shortcomings in the strategy for forming the training sample can lead to an explosive increase in computational costs during training. This effect is called the "curse of dimension". The purpose of this work is to describe the methods of compression of the training sample, based on the identification of boundary precedents of the training sample and getting rid of the ballast of internal precedents that do not take part in decision making. The paper discusses various approaches to managing the size of the training sample. It is shown that the method of constructing a hull, bordering the feature space of the Smart Protection Device, based on the use of alpha forms, allows compressing the training sample without compromising its information capacity. The application of the compression of the training sample of the Intelligent Discriminator of the modes of ground short circuits in the recognition of a single-phase short circuit to ground is demonstrated. The training results confirm that the training set retained its effectiveness even after compression.