Adaptive, Predictive Controller for Optimal Process Control

S.K. Brown, C.C. Baum, P.S. Bowling, K.L. Buescher,
V.M. Hanagandi, R.F. Hinde, Jr., R.D. Jones, W.J. Parkinson
Los Alamos National Laboratory, Los Alamos NM, USA


One can derive a model for use in a Model Predictive Controller (MPC) from first principles or from experimental data. Until recently, both methods failed for all but the simplest processes. First principles are almost always incomplete and fitting to experimental data fails for dimensions greater than one as well as for non-linear cases. Several authors[1] have suggested the use of a neural network to fit the experimental data to a multi-dimensional and/or non-linear model. Most networks, however, use simple sigmoid functions and backpropagation for fitting. Training of these networks generally requires large amounts of data and, consequently, very long training times.

In 1993 we reported on the tuning and optimization of a negative ion source using a special neural network[2]. One of the properties of this network (CNLSnet), a modified radial basis function network, is that it is able to fit data with few basis functions. Another is that its training is linear resulting in guaranteed convergence and rapid training. We found the training to be rapid enough to support real-time control.

This work has been extended to incorporate this network into an MPC using the model built by the network for predictive control. This controller has shown some remarkable capabilities in such non-linear applications as continuous stirred exothermic tank reactors and high-purity fractional distillation columns[3]. The controller is able not only to build an appropriate model from operating data but also to train the network continuously so that the model adapts to changing plant conditions.

The controller is discussed as well as its possible use in various of the difficult control problems that face this accelerator community.

[1] P.J. Werbos, T. McAvoy, and T. Su, Handbook of Intelligent Control (Van Nostrand Reinhold, New York, NY, 1992) 286.
[2] W.C. Mead, S.K. Brown, R.D. Jones, P.S. Bowling, C.W. Barnes, "Adaptive Optimization and Control Using Neural Networks," Proceedings of the 1993 ICALEPCS Conference, Berlin Germany, 1993, 309-315.
[3] C.C. Baum, K.L. Buescher, R.D. Jones, W.J. Parkinson, and M.J. Schmitt, "Two-timescale, Model Predictive Control of Two Simulated Chemical Plants: Lagged-CSTR and Distillation Column," Los Alamos Report (LA-UR-94-1039).