Neural control using a hierarchically performed BP algorithm

Deals with the design of a neural network regulator (NNR) for nonlinear industrial dynamic systems, based on the multilayer perceptron (MLP) and using a hierarchically performed backpropagation (BP) algorithm. A novel network architecture is employed for regulator design: the NNR consists of two subnetworks, one of which is used for I-O (input-output) mapping, while the other acts as an adaptive controller. The BP algorithm is employed to reproduce a nonlinear relation between the inputs and outputs of the plant and to update regulator parameters. The proposed architecture has the flexibility for adding more sensory information and facilitates extension to multi-input, multi-output systems and multivariable controllers. The operation of the NNR does not require a reference model or inverse system model, or any probing signals, and can produce more acceptable control signals than are obtained using the sign of the plant errors during the backpropagation procedure. The regulator has been applied to a complex nonlinear turbogenerator system