Detecção e classificação de distúrbios de tensão usando redes neurais ARTMAP Euclidiana modificada com treinamento continuado

This research presents a methodology to detect and classify voltage disturbances in electrical distribution systems using as tool, statistical and multilevel direct analysis of voltage signals in wavelet domain for the detection process and modified Euclidean ARTMAP neural network with continuous training for the classification (ARTMAPMECT). These two tools provide velocity, security and reliability for decision making when it is necessary to restore the normal operations. The ARTMAP-MECT is able to adapt to continuous changes at the electrical system, e.g. to provide a diagnosis of unknown disturbances by the neural network Furthermore, it allows the inclusion of new patterns in the memory of the neural network without reinitializing the training, which is not possible in other neural networks. It is a new approach of neural network application where the knowledge is permanently extracted from the system under analysis from information available in the system operation (results from analysis, simulation, etc.). The voltage signals that are necessary to execute the voltage disturbance diagnosis are obtained from simulations of a radial feeder using the ATP (Alternative Transients Program) and MATLAB using the discrete wavelet transform, multiresolution analysis and energy concept. The obtained results show the efficiency of the proposed methodology allowing the application in distribution protection, aiding the operators in decision procedures to restore the system and contributing to improve the power quality provided to the users.