(247d) Exploiting Connectivity Structure for Online Selection of Primary Controlled Variables

While there are numerous works on controller design, research in the area of optimal control structure design is still very limited. Optimal selection of the primary controlled variables by considering plant economics as well as controllability is critical for optimal control structure design. However, due to change in the operating conditions, control objectives, as well as operating constraints, optimal primary controlled variables can change over time. Due to large number of candidate variables that lead to significant number of combinatorial possibilities even for small plants, online selection of primary controlled variables is computationally prohibitive. In this work, we propose a novel approach that exploits the structural connectivity of the plant to decompose the large-scale problem into smaller subsets. This approach makes parallel solution of the controlled variable selection problem feasible thus achieving solution within the target execution time.

For determining the changing connectivity among plant sections, we draw from the neuroscience literature where the connectivity among cortical/sub-cortical areas are determined using the fMRI responses. At a cortical level, the neuronal populations can be modelled as states that dynamically evoke brain responses as a function of inputs. Thus, by modeling islands of sub processes in a chemical plant as cortical/sub-cortical areas, the effective connectivity of these islands of sub processes can be extracted using bilinear approximation. These connectivity matrices can then be employed as a basis for controlled variable selection.

In this approach, a bilinear model of the process is developed where the model parameters represent intrinsic coupling among the states, describe the influence of extrinsic inputs on the states and as well as capture the effect of inputs on coupling. These parameters are identified using a Bayesian framework. The algorithms developed are implemented on a model of an integrated gasification combined cycle plant with CO₂ capture. This plant includes significant mass and energy interactions with strong change in the coupling as the inputs are changed making it a prefect test case for the developed algorithm. Using an expectation maximization algorithm with uninformed priors, the algorithm provides information about the connectivity strengths among various plant sections. It is observed that the algorithm correctly captures the â??trueâ?? connectivity as would be expected from the first-principles model, plant configuration and operating conditions. The presentation will include algorithmic details as well as detailed results that provide insights into change in connectivity induced by external disturbances.