Active power filters compensate for harmonic distortion created by nonlinear loads. Compensation may not be complete due to errors or other system influences. A neural network active filter controller is discussed in this paper which learns to improve the compensation of the harmonic distortion. The aim of the controller is to reduce either the amount of utility current or voltage harmonic distortion. A comparison between frequency and time domain neural network controllers based on steady state and transient performance is considered. Improvements in the transient response can be made by using a neural network and a traditional active filter controller combination. The hardware and software requirements of implementing the time and frequency domain neural network controllers are also discussed.