Dynamic real-time optimization and control of an integrated plant

Applications of the existing steady-state plant-wide optimization and the single-scale dynamic optimization strategies to an integrated plant with a material recycle loop have been impeded by several factors. While the steady-state optimization strategy is very simple to perform, the very long transient dynamics of an integrated plant have limited the execution rate of the optimizer to be extremely low, yielding a suboptimal performance. On the other hand, the single-scale dynamic plant-wide optimizer that executes at the same rate as local controllers would require an exorbitant on-line computational load. In addition, it may be sensitive to high-frequency dynamics that are not relevant to the interaction dynamics of the plant, which are slow-scale in nature. This thesis presents a novel multi-scale plant-wide optimization strategy suitable for an integrated plant with recycle. The dynamic plant-wide optimizer in this framework executes at a slow rate in order to track the slow changes that are relevant to the plant-wide interactions and economics, while leaving the fast changes in unit operations to be handled by local controllers. Moreover, the computational requirement of solving the optimization problem will be much smaller than that of the single-scale dynamic optimizer running at a very high rate.; An important issue of the suggested method is in obtaining a suitable dynamic model for optimization. When dynamic first-principles models are available, model reduction techniques that reduce model order, while retaining slow-scale information in the frequency range of interest by the optimizer can be used. On the other hand, when fundamental constitutive equations are not available, system identification experiment needs to be performed to obtain information on the interaction dynamics of the system. The difficulties in this process are how to design input signals to excite this ill-conditioned system properly and how to handle the lack of slow-scale dynamic data when plant experiments cannot be conducted for a very long period of time compared to the plant's settling time. This work addresses the experimental design and suggests a new grey-box modeling method to incorporate steady-state information to improve model prediction quality.; To extend this framework to a nonlinear integrated plant while ensuring a small on-line computational requirement and robustness against uncertainties, the Approximate Dynamic Programming (ADP) framework is adopted. This method offers advantages over conventional mathematical programming based approaches in that it can compute an optimal operating policy under uncertainties off-line. The on-line multi-stage optimization problem can be reduced to a single-stage problem, thus requiring much less real-time computational effort. In process system community, where the system has continuous state and action space, a simulation-based ADP method coupled with a function approximation scheme has been proposed. However, the existing ADP framework is inadequate to handle an integrated plant problem, which has a large action space and a high-dimensional system model. In this thesis, we use a case study to show the drawbacks of the existing mathematical programming framework and motivate the ADP approach. We combine a local gradient search technique and a nonlinear model reduction approach to overcome a very large off-line computational requirement of the existing ADP approach. The resulting framework shows superior performance in solving an optimal control problem of an integrated plant in both deterministic and stochastic cases and can be generalized to larger problems.