Abstract:
Traditionally the tuning of DMC-type multivariable controllers is done by trial and error. The APC engineer would choose arbitrary starting values and test the performance on a simulated controller. The engineer would then either increase the values to suppress movement more, or decrease them to have the manipulated variables move faster. When the controller performs acceptably in simulation, then the tuning is improved during the commissioning of the controller on the plant. This is a time consuming and unscientific exercise and therefore often does not get the required attention, leading to unacceptable controller behaviour during commissioning and sub-optimal control once commissioning is completed. This dissertation presents a new method to obtain move suppression factors for DMC type multivariable controllers. The challenge in choosing move suppressions lies in the multivariable nature of the controller. Changing the move suppression on one manipulated variable will not only change the performance of that manipulated variable, it will also change the performance of every other manipulated variable with models to the same controlled variables. In the same way, changing the steady state cost of a manipulated variable or the equal concern error of a controlled variable will also affect the behaviour of every other manipulated variable with shared models. There have been attempts to calculate the required move suppression factors mathematically. Some methods used an approach that is based on the premise that move suppression factors that present a well-conditioned controller matrix will provide a well behaved controller in terms of tuning. Some other methods focussed on providing parameters that will cause desirable controlled variable response, either by determining tuning parameters offline, or by re-tuning the controller in real time. The method described in this paper uses a Nelder Mead (Nelder and Mead, 1965) search algorithm to search for move suppressions that will provide acceptable control behaviour. Acceptable behaviour is defined by characterising the dynamic move plan calculated by the controller for each of the manipulated variables, or by characterising the controlled variable path that will result from the manipulated variable moves. The search algorithm can change the move suppressions, the steady state costs, or the move suppression multipliers as used in DMC type controllers. Copyright
