ON-LINE TRAINING OF ARTIFICIAL NEURAL NETWORKS FOR CONTROL OF A HEAT EXCHANGER TEST FACILITY

摘要

Artificial neural networks (ANNs) are known to be capable of simulating steady-state and transient thermal processes. They have also been used with control purposes in thermal systems such as heat exchangers. Before ANNs can be used to model or control a system they need to be trained and this process is usually performed off-line. It is known that the characteristics of thermal components such as cross flow heat exchangers vary with respect to time mainly due to fouling effects. In this work, the excellent adaptive characteristics of ANNs are used to modify the weights and biases of the network to learn the new characteristics of a thermal system. We control the outlet air temperature of a heat exchanger test facility by means of varying the mass flow rate of the external or internal fluid. The parameters of the ANNs are modified depending on the error obtained between the desired outlet air temperature and its measured value. Since it has been shown that neurocontrollers can become unstable, the weights and biases of the ANNs are also modified considering the stability conditions of the closed loop system considered as a nonlinear iterated map. An internal model control scheme is used together with an integrator to perform the controlling action. We also take advantage of being able to vary the internal and external mass flow rates to implement the minimization of a performance index that quantifies the energy consumption. It is shown numerically and experimentally that the neural network is able to control the thermal facility, and is also able to adapt to different disturbances applied to the system, while minimizing the amount of energy used.