The main objective of this research is to develop a method to construct a neural tuner of PI-controller parameters. Such tuning is necessary to improve energy efficiency of nonlinear industrial plants since the PI-controller is linear, and its parameters certain values are optimal only for a certain plant functioning mode. The tuner consists of a neural network calculating the controller parameters values and a rule base scheduling moments of the network training and learning rates values. Methods to select both the neural tuner structure and an appropriate neural network structure are proposed. The general network structure is defined by Cybenko, Horknick and Gorban theorems. The number of input layer neurons is chosen in accordance with a PI-controller difference equation. The number of neurons in a hidden layer calculation method is based on the number of input layer neurons, delayed setpoint and plant output signals used in the input layer and the number of points, which are used to average network input signals. Such methods do not use the plant model. Stability of a control system with the tuner is estimated for each transient using technical sustainability criteria. Proposed methods are applied to control empty and loaded electroheating furnace. Obtained results show that the tuner usage allows to achieve up to 6% plant energy efficiency improvement in comparison with the conventional PI-controller with constant parameters usage.