SMART ALGORITHM FOR INTERVAL ESTIMATION OF ARC-QUENCHING REACTOR PARAMETERS
The purpose of the arc suppression reactor is to reduce the capacitive current of the network to a safe level where the single-phase earth fault current at the fault location does not exceed five amperes. The current reduced to a permissible level prevents open arcing at the fault location. For proper operation of this device, the arc suppression reactor control automatics needs to adjust the zero-sequence circuit to resonance, which balances the capacitive current of the mains and the inductive current of the reactor. To perform this tuning, it is not necessary to have information about the absolute values of the parameters of the zero-sequence circuit, but by determining them, the automation device is able to solve a wider range of tasks related to network diagnostics and increasing the efficiency of the arc suppression reactor. In this article we consider an approach to solving the problem of parametric identification of arc suppression reactor using the method of interval estimation of object parameters. The information about the operation modes of the arc suppression reactor is obtained by means of a simulation model of the object. Using the observed values, the object parameters are obtained by use of the inverse function to the simulation model. The dependence of the object parameters on the observed parameters is approximated using upper and lower parameter estimation models. The quantile regression method was applied to tune the estimation models. The need to increase the generalization ability of the algorithm is revealed. The method of adjustment of parameters of regularization of learning process to increase generalization ability of algorithm without increase of informativity of data in a training sample is offered. The results of algorithm performance are presented on the example of magnetization branch parameters estimation of arc suppression reactor. The boundaries of the interval of equivalent magnetic core loss resistance and magnetizing inductance are obtained. The limitations of the methods are analyzed, and recommendations for improving the quality of the algorithms are given.