An on-line trained neural network with an adaptive learning rate for a wide range of power electronic applications

Artificial neural networks (ANNs) are particularly useful to represent the input-output relationships of nonlinear time-varying systems; such applications in power electronics and adjustable speed drives have been reported in the recent literature. Continuous online training of such systems requires high speed signal processing. Commercially available ANN hardware is too slow for fast power electronic systems. This paper proposes a new fast online random weight change training algorithm which uses an adaptive learning rate and is suitable for very high speed VLSI implementation. It requires little or no input from the user and is self-commissioning