Nonparametric time-varying phasor estimation using neural networks

A new approach to nonparametric signal modelling techniques for tracking time-varying phasors of voltage and current in power systems is investigated. A first order polynomial is used to approximate these signals locally on a sliding window of fixed length. Non-quadratic methods to fit the linear function to the data, give superior performance over least squares methods in terms of accuracy. But these nonquadratic methods are iterative procedures and are much slower than the least squares method. A neural network is then used to model the non-quadratic methods. Once the neural network is trained, it is much faster than the least squares and the non-quadratic methods. The paper concludes with the presentation of the representative testing results.