Adaptive Tuning – Model Switching

During the development of the DeltaV adaptive tuning capability, we researched and evaluated many of the most common techniques for implementing adaptive tuning. Also, as part of this research we came across a number of technical articles on a relative new adaptive tuning technique known as model switching. This technique depends on the evaluation of a bank of models to determine which model most accurately matched the process behavior. To avoid the memory and processor limitations associated with such an approach, Willy Wojsznis proposed that adaptation be performed through model-parameters interpolation. This approach proved to be quite effective and drastically reduced the number of models that must be evaluated. As described in Chapter 5 of Advanced Control Foundation – Tools, Techniques and Applications, a PID adaptive controller with model switching and parameter interpolation may be illustrated as shown below

The process identification is based on the evaluation of multiple models. Each model consists of three parameters i.e., gain, deadtime, lag. Improved convergence and a reduction in the number of models is achieved by performing parameter adaptation sequentially, one parameter at a time. In this way, the number of model combinations for the first order-plus-deadtime model has been reduced to 3×3 = 9. Also, faster convergence is achieved by using the original data set and performing adaptation iteratively by running the algorithm several times. In an iterative adaptation procedure in which one parameter is updated over a calculation cycle, the updating is performed in a sequence of, first, process gain, then deadtime, and, finally, time constant over the model adaptation cycle as illustrated below.

After each iteration, the parameter adaptation range is made smaller, improving the accuracy of the parameter identification. Several criteria for comparative model validation as well as statistical model evaluation techniques are used to provide a reliable means for model validation. After model adaptation completes, controller redesign begins using the first order-plus-deadtime process model. Any model-based tuning rule the user chooses can be applied.