Control relevant long-range plant identification using recurrent neural networks

In adaptive critic control systems design as well as other control systems design schemes, e.g., model-based predictive control, the plant model has to be iterated to predict many time steps ahead into the future. A commonly used implementation is to employ a parallel identification structure. It has been justified for linear estimation that (assume that estimates to the real plant are biased) long-range prediction models are less sensitive to high frequency noise, whether actual noise or caused by model-plant mismatch. We address feasibilities of using recurrent neural networks for long-range plant identification. We examine the existence, training and performance of such recurrent neural network identifiers.