Classification of Power Quality disturbances using Multi-Valued Neural Networks and Convolutional Neural Networks

This paper proposes two monitoring methods based on different machine learning techniques capable of detecting electrical disturbances in low voltage power grids. Nowadays, the massive use of electronic loads and controllers introduces several distortions in the electrical distribution grid due to the non-linear behavior of these devices. The detection and classification of disturbances play a fundamental role in cost reduction for users and overall energy quality improvement. Machine learning algorithms are currently being exploited to achieve this goal. The paper presents a comparison between a classifier based on Convolutional Neural Networks and a Multilayer Neural Network with multi-Valued Neurons. In both cases the inputs of the classification tool are time-domain samples of the phase voltage extracted from an Italian distribution grid at 230/400 V. The training phase of the classifiers is carried out in the Matlab environment using the analytical form of five disturbances: voltage sag, voltage swell, harmonic distortion, voltage notch and interruption. The performances of the monitoring systems are compared in terms of classification rate through the training procedure followed by testing using the data from real measurements. Both classifiers have an accuracy of about 90% during the training phase and comparable performance with measured data.