Deep Learning in Relay Protection of Digital Power Industry

Abstract

Although modern relay protection exhibits all properties of an intelligent system, it has not yet fully acquired abilities to learn, adapt and recognize the modes of the protected electrical network. To give it these advantages, it is necessary to solve the central problem which is to distinguish the areas controlled by the protection of the electrical network modes parameters. In relay protection all modes of network are divided into watched and alternative. In the first mode, the protection should be tripped, and in the second - the tripping is strictly prohibited. The problem of ensuring selectivity of protection can be considered at the rate of process as the establishment of belonging of the arriving data of an electrical network mode to a certain class in space of controlled parameters, i.e. having determined them to a class of the watched or alternative modes. Traditional methods of relay protection learning are based on the application of the characteristics of operation differentiating the watched and alternative modes, thereby revealing some similarities with elements of the theory of artificial intelligence. At the same time the problem of finding acceptable operation border is solved with grace by methods of machine learning. Despite external similarity of schemes of algorithms, their main difference consists in a way of a task of characteristic of operation of protection. In traditional relay protection characteristic of operation is stored in the permanent memory, and in the protection with artificial intelligence - in the rewritable memory and is a part of neurons. The paper presents a solution to the problem of differentiation of the electrical network modes on the basis of deep learning, considering the problem of formation of the tripping areas, for example, resistance relays as a definition of belonging to the relay measurements to certain classes in the space of controlled parameters. Algorithms of machine learning have universality, efficiency and allow approaching scrupulously the choice of characteristics of operation, using all the arsenal of intelligent classifier. As a result, intellectual relay protection gains ability to differentiate difficult untied areas of precedents (the modes of an electrical network) containing enclaves of the alternative modes. Besides, intellectual relay protection has a possibility to correct characteristic of operation in the conditions of its operation by training of neural network at new precedents. However, for this purpose the operational personnel have to give signs to new data, turning them into precedents. Thus, intellectual relay protection has ability to adapt to changes of an electrical network. The paper covers mathematical foundations of the precedents sets separation of different modes of the electrical network on the example of the intelligent resistance relay using a support vector machine method. The advantage of the method consists in using the uniform principles to classify the modes of an electrical network both in case of linear, and in case of nonlinear separation of precedents. At the same time the solution of the linear separability problem is considered as a solution of the quadratic programming problem in the traditions of the theorem of Karush-Kuhn-Tucker. If it is impossible to recognize the parameters of a mode by a linear classifier, a nonlinear classifier is used, applying the mapping of the initial case space using special kernels to a higher-dimensional straightening space where the set of the modes becomes again linearly separable. Possibilities of application of a support vector machine method to solve the problem of classification and deep learning of relay protection are shown. On the example of an intelligent resistance relay, the mechanism of the tripping characteristic adaptation to changes in network parameters is illustrated.