Identification of types of distortion sources in power systems by applying neural networks

To aid in the task of locating the geographical origin of distortions in power systems, the population of devices that causes distorted or non-sinusoidal conditions was divided into 3 main categories: Power electronic, arcing and ferromagnetic devices. An artificial neural network was developed that categorises randomly picked devices accordingly. The magnitude of the current harmonics was identified as the best distinguishing features. Typical relative current harmonic values for each device were found. Each category of distortions was investigated in terms of its harmonic content. It was found that significant correlations exist between the individual current harmonics of the same device. Matlab simulations indicated that a three layer neural network with five hidden units optimised by the Bayesian regularisation algorithm, provided the best results in terms of speed and accuracy.