Dynamic neural networks for inverse dynamics based control of evaporator

Abstract

This paper compares inverse dynamics identification for controlling an evaporator using dynamic neural networks developed in two different control strategies, namely, conventional direct mapping NN (DMNN) with RBF nonlinear static functions and RBF dynamic neural unit (DNU) neuronal models as perceptrons. In spite of their success, DMNN suffered from the problem of curse of dimensionality which involves parametric dimensionality and structural dimensionality. Hence this paper gives a novel NN architecture consisting of simple nonlinear dynamic blocks, termed DNUs as perceptrons to overcome this major problem and a comparison is done with a conventional DMNN to validate the proposed method. In the experimental plant, the evaporator heat flow rate and secondary fluid outlet temperature are to be controlled while keeping refrigerant superheat temperature in the range 4-7 K at the evaporator outlet by manipulating refrigerant and evaporator secondary fluid flow rates. Therefore a multi-input multi-output controller is required for its proper control. The effectiveness of the proposed novel dynamic NN controller is demonstrated through reduced number of activation functions with lesser calculation time and efficient error convergence in training. Again, the probability of error convergence to a global minimum is quite high when NN structure gets simple. Therefore, this kind of dynamic NN can handle real-world applications efficiently. The inverse dynamics identification was elaborated using experimental data from the ammonia refrigerant evaporator and the proposed NN architecture assures promising results.