Architecture and training algorithm of the adjusting neural network for accurate model tuning

Abstract

Mathematical models used for control, such as for generating control references, have to be identical to the actual controlled objects to improve control accuracy. For this, their parameters are often tuned by model errors derived from the detected outputs of the controlled objects and calculated ones using the mathematical models. Therefore accurate identification between the model errors and appropriate tuning values of the parameters is necessary to complete tuning with small tuning numbers and no steady state errors. In this paper, we propose The Adjusting Neural Network (AJNN), which is the extended model of conventional Multi-layered Neural Networks (CNN), to identify the non-linear relationship between them accurately. AJNN consists of two Multi-layered Neural Networks, namely, the Error Calculation Neural Network (ECNN) is added in parallel to CNN, where the ECNN calculates the output error of CNN and subtracts it from the output of CNN, to obtain accurate tuning values. Training method for AJNN is also proposed, where modified Back-propagation developed for reducing the error of AJNN and the conventional Back-propagation for decreasing the output of ECNN to zero, are applied to AJNN alternately. Finally, AJNN is evaluated with the model tuning of temperature control for reheating furnace plants and demonstrated to be effective to improve the accuracy of tuning and decrease tuning numbers. © 1995, The Institute of Electrical Engineers of Japan. All rights reserved.