Estimação de componentes harmônicos em sistemas elétricos de potência utilizando redes neurais artificiais tensoriais

Abstract

The Widrow-Hoff rule is the simplest and most used method to update the synaptic weights of an artificial neural network. This is due to its simplicity of application and parametric robustness, properties of interest in several applications. However, in struc tures with high numbers of parameters, this algorithm has high computational cost and low speed of convergence. In this work, a new rule for updating the synaptic weights of a neural network is presented. The proposed method uses the concepts of linear opera tor separability, a typical property of tensors. Once this property is given, a stochastic gradient algorithm can be derived with a significant increase in the learning speed and a reduction in the computational cost when compared to the Widrow-Hoff algorithm. As the number of neural network parameters increases, the performance difference between the two algorithms is enhanced. Another relevant property of the proposed method is that its learning speed is inversely proportional to the computational cost. Due to the great impact generated by the growth of non-linear loads in the electrical power system, which cause distortions in the voltage and current waveforms, the proposed methodology is applied in harmonic estimation problems.