469943 Utilizing Memory in Process Control
In this work, a dynamic sparse coding (DSC) algorithm is developed for nonlinear time-varying systems for the purpose of using the individual sparse coefficients as classifiers. The algorithm developed here allows for evaluation of a linear combination of a subset of basis functions drawn from a candidate set by using an active-set likelihood maximization algorithm. The dynamic data are preprocessed and then used for obtaining the dynamically produced representations. The identification procedure is solved by an efficient sparse coding formulation using the Bayesian (Schwarz) Information Criterion (BIC) as a method for determining the minimal number of basis functions to fully describe the system being considered. This sparse representation is then used to infer the similarity of the control problem. Finally, if there is acceptable similarity to a past control problem, then the experience from the past controller performance is exploited to obtain improved performance by implementing a weighted scheme determined based on the similarity ratio. The sparse coding representation is also utilized for dynamically adapting the process model that is used in a model-based control technique. The sparse coding problem is formulated to ensure that if the process remains time-invariant, the sparse representation remains relatively invariant to changes in the inputs.
This approach is validated on an acid gas removal unit as part of an integrated gasification combined cycle plant. This unit consists of large number of equipment items, is highly nonlinear, and involves significant mass and heat integration along with considerable time delay. It is observed that even if the controller performance is poor to start with, repeated use of past experiences leads to almost perfect control.