A predictive control scheme based on neural networks

Purpose - To develop a new predictive control scheme based on neural networks for linear and non-linear dynamical systems. Design/methodology/approach - The approach relies on three different multilayer neural networks using input-output information with delays. One NN is used to identify the process under control, the other is used to predict the future values of the control error and finally the third one is used to compute the magnitude of the control input to be applied to the plant. Findings - This scheme has been tested by controlling discrete-time SISO and MBIO processes already known in the control literature and the results have been compared with other control approaches with no predictive effects. Transient behavior of the new algorithm, as well as the steady state one, are observed and analyzed in each case studied. Also, online and offline neural network training are compared for the proposed scheme. Research limitations/implications - The theoretical proof of stability of the proposed scheme still remains to be studied. Conditions under which non-linear plants together with the proposed controller present a stable behavior have to be derived. Practical implications - The main advantage of the proposed method is that the predictive effect allows to suitable control complex non-linear process, eliminating oscillations during the transient response. This will be useful for control engineers to control complex industrial plants. Originality/value - This general approach is based on predicting the future control errors through a predictive neural network, taking advantage of the NN characteristics to approximate any kind of relationship. The advantage of this predictive scheme is that the knowledge of the future reference values is not needed, since the information used to train the predictive NN is based on present and past values of the control error. Since the plant parameters are unknown, the identification NN is used to back-propagate the control error from the output of the plant to the output of the controller. The weights of the controller NN are adjusted so that the present and future values of the control error are minimized