Implementation and performance evaluation of an on-line neural network control scheme

This paper investigates the ability of the multi-layer perceptron (MLP) neural network to accurately model a laboratory scale non-linear process, liquid level control. The neural network model of the process is then included in an on-line predictive control strategy where the performance of the controller is evaluated when both single and multi-step-ahead predictions are used in the control algorithm. A method of conditioning the process input-output data before it is presented to the neural network is described It has been found that using the described coding rather than a frequently used method of conditioning the process data results in the neural network being able to predict more accurately when being used recursively. The performance of the neural network based model scheme is also compared to that of a well tuned PI controller. Finally an insight into the stability of the closed loop neural control scheme is examined in a novel application of system identification techniques.