Swarm-intelligently trained neural network for power transformer protection

Abstract

The paper presents the particle swarm optimization technique (PSO) to train multi layer neural network for discrimination between magnetizing inrush current and internal fault current in power transformers. The discrimination process relies on the harmonic components of both inrush and fault currents. The features were extracted from the wave of the differential current by using the fast Fourier transform (FFT). The output of the neural network is trained to respond "high" or "1" for fault current (trip command of the differential relay) and "low" or "0" for inrush current (no trip command). Compared to the back propagation (BP) training method, neural networks using the particle swarm optimization technique is more accurate (in terms of sum square errors) and also faster (in terms of number of iterations).