Neural network control of a nonlinear dynamic plant with a predictive model

The paper considered the possibilities of applications of neural network technologies to control a dynamic plant with nonlinear properties. To give the control system the desired dynamic property, the use of a neural network predictive controller is proposed. The model of the control plant is in the form of a multilayer forward-directional neural network, which allows us to construct a controller using generalized equation methods with prediction. A neural network control algorithm with prediction based on minimizing the quadratic quality functional is proposed. The algorithm makes it possible to minimize the root mean square error of regulation and the control signal rate of change. To determine the sequences of optimal control impacts, the application of the Newton-Raphson method is proposed. To reduce computational costs when receiving control signals, the decomposition of the original matrix, represented as a Hess matrix, is carried out. To predict the behavior of a control plant, a formula is proposed for calculating the gradient of a neural network, discrepant by the possibility of its use in the real-time mode of the control. The proposed algorithm of the neural network control with predictive allows higher quality control of complex nonlinear dynamic plants in the real-time mode.