Neural network structure selection method to solve linear controllers parameters adjustment problem

The problem of a neural network structure selection is considered. It is used as a part of a neural tuner to adjust P-, PI- or PID-controllers parameters online to control nonlinear plants. A method to find a number of network layers, neurons in each of them, choose activation functions and calculate delay time for network delayed inputs is proposed. Such method does not need a plant model. An algorithm to synthesize and initialize the neural network for the neural tuner with the help of a-priori known data about the plant is developed. Having made the experiments with plant models and two electroheating furnaces, we conclude that neural tuner helps to achieve both time and energy consumption decrease to complete setpoint schedule in comparison with conventional linear controller. This fact shows that proposed method is valid.