Dynamic Causal Modelling and Its Application to an Acid Gas Removal Unit
Temitayo Bankole, Debangsu Bhattacharyya
Department of Chemical Engineering, West Virginia University, Morgantown, WV, USA
Controlled variables are typically selected ad hoc without due consideration of their economic and control performances. Recently our group developed a thorough and systematic approach to controlled variable selection (Jones, et al., 2014) by considering tradeoff between economic performance and control performance. However as the process model changes during operation, the optimal controlled variables can change. Due to the large number of candidate variables and the large scale optimization problem that needs to be solved for optimal controlled variable selection, it is computationally intractable to perform the selection process dynamically for a large scale process. One option is to divide the plant sections based on their connectivity information and then perform the selection process separately for sections that are weakly connected. However, this will require determining how the connectivity information between different plant sections changes dynamically. Drawing from the neuroscience literature (Friston, 2003) that focuses on dynamic changes in the connectivity among cortical and sub-cortical areas in the human brain, a dynamic causal model is developed where the dynamic data collected from the system are used to determine how the latent connectivity and the induced connectivity change due to external and internal stimulations.
At a cortical level, the neuronal functions including their interactions can be modelled as states that dynamically evoke brain responses as a function of inputs. This is, in essence, similar to a chemical plant where the connectivity strength changes depending on the state of the system and internal and external perturbations. Thus the quantitative information about effective connectivity of the plant sections can be extracted using a bilinear approximation by using the transient data. These connectivity matrices can then be employed as a basis for selecting the controlled variables dynamically. The parameters reflecting the connectivity information are obtained by using an expectation maximization algorithm based on a Bayesian approach. Finally, the algorithm is deployed to a Selexol-based acid gas removal (AGR) unit for selective capture of CO2 and H2S as part of an integrated gasification combined cycle (IGCC) unit. The AGR process is highly nonlinear with strong mass and heat integration among different process sections. Our work shows that it is possible to capture the dynamic changes in the connectivity matrix by using appropriate measurement data. In this presentation, we will also discuss the computational issues and efficiency of the proposed algorithm.
Friston, K. J. L. H. W. P., 2003. Dynamic causal modelling. Neuroimage, Issue 19, pp. 1273-1302.
Jones, D., Bhattacharyya, D., Turton, R. & Zitney, &., 2014. Plant-wide control system design: Primary controlled variable selection. Computers & Chemical Engineering, pp. 220-234.
See more of this Group/Topical: Computing and Systems Technology Division