Neuro Algorithm Accuracy Evaluation for The Anomalies Detecting in Overcurrent Protection Operation

To control information exchange in the digital substation, it is possible to use machine learning methods, the artificial neural networks (ANN), in particular. The ANN training, as a multifactor optimization problem, and its testing make it possible to detect anomalies in information flows and operating modes of the secondary equipment in electric power industries. Using examples for neuro-modeling of overcurrent protection operation mode in a three-phase network, the paper shows the features of a neural network functioning based on an elementary perceptron and estimates the accuracy of the corresponding models. The structure of the simplest neural network is optimized and accuracy evaluation of the neural network algorithm is presented depending on the training sample volume (from 1000 to 50000 records) and the number of training epochs. It was found that neuromodelling makes it possible to evaluate with a high accuracy the value for overcurrent protection adequate trip setting. The deviation range of current values in each phase, when the neural network makes errors in data recognition, does not exceed one percent of the threshold value., The neural network in this range is very "sensitive" to changes in input signals. It is shown that if emergency modes with adequate actuation of relay protection are considered as standard secondary equipment operation modes, then neuromodeling and corresponding signal processing make it possible to detect anomalies in the standard secondary equipment operation modes in the electric power system.