Predictability of back propagation and discrete Hopfield neural networks in harmonic compensation systems

A quality tool for predicting control signals is indispensable for effective harmonic compensation in AC power distribution systems. Notably in the literature, either back propagation (BP) or Hopfield neural networks (HNN) has claimed to provide quality control signals to achieve the desired harmonic reduction results. This paper evaluates the predictability of BP and HNN in terms of convergence behaviour and learning capability, as applied to the reduction of load generated current harmonics in a variable speed DC drive. Using the same real current harmonic data, our test results confirm that BP has a larger dynamic harmonic range whereas discrete HNN, due to its interconnection structure, needs larger size of memory map.