Neural Network Predictive Control

In the first approach, direct substitution of ANN model in GPC is used. This is based on idea that GPC is not related to any particular process model, but it rather needs only two process characteristics: step and free responses. These are obtained from ANN.

Next, a more complicated scheme is used where one ANN is used as a model while the second forms predictive controller - it is learned on-line. Thus the optimised variables are not future control increments as usual, but ANN weights. After some time, the controller ANN is able to control on-line without further learning. To make on-line learning possible, stochastic approximation is used in optimisation.

Neural networks are also used in predictive control where the optimiser is based on Iterative Dynamic Programming (IDP) - a method for dynamic optimisation. The whole is used in the receding horizon framework. As ANN are fast predictors, on-line implementation of the algorithm is possible.
 

Journals

• Rusnák, A., Fikar, M., Latifi, M. A., Mészáros, A.: Receding Horizon Iterative Dynamic Programming with Discrete Time Models. Computers chem. Engng., 25(1), 161-167, 2001. Abstract
• Rusnák, A., Fikar, M., Mészáros, A.: Receding Horizon Control Using Modified Iterative Dynamic Programming and Neural Network Models. Computers chem. Engng., 23, S297-S300, 1999. Abstract
• Najim, K., Rusnák, A., Mészáros, A., Fikar, M.: Constrained Long-Range Predictive Control Based on Artificial Neural Networks. Int. J. Systems Science, 28 (12), 1211 - 1226, 1997. Abstract
• Rusnák, A., Fikar, M., Najim, K., Mészáros, A.: Generalized Predictive Control Based on Neural Networks. Neural Processing Letters, 4(2), 107-112, 1996. Abstract

Conferences

• Rusnák, A., Fikar, M., Mészáros, A.: Receding Horizon Control Using Modified Iterative Dynamic Programming and Neural Network Models. Computers chem. Engng. / ESCAPE 9, 23, S297-S300, 1999. Abstract, Poster.PDF