Modeling Updating in Real-Time Optimization

A model-based real-time optimization (RTO) system can improve the operating profit of the plant by tracking the changing optimum responding to distubances. Naturally, the RTO tracking capability relies on the accuracy of the model and the estimated parameters. This thesis develops the technologies to enhance RTO performance by acquiring better data sets for updating and investigates the effect of model fidelity on RTO performance. The main contribution of this thesis is to improve updating through the use of multiple data sets (current and recent past data sets), including the possibility of judiciously designed plant experiments. A moving window approach is used to keep a record of current and historical data sets for updating. When plant variation exists in the data sets, extra parameters, which require plant variation for accurate estimation, can be updated to reduce the plant/model mismatch. The updater diagnostic tests are implemented to identify the maximum number of parameters that can be estimated from the available data sets. The parameter estimation problem is formulated by incorporating prior knowledge of disturbance frequency to track the disturbances with different dynamics. Accurate estimation of some of the adjustable parameters requires plant variation which can be generated by disturbances. When the data sets do not have sufficient plant variation for updating, limited plant perturbation can improve updating. A new expected profit approach for designing plant experiements is developed and integrated with the RTO system. Experiments are designed to maximize the overall expected profit to trade off the benefits achieved by improved model and cost of experimentation. The effect of model fidelity on RTO performance is investigated using several case studies including an industrial boiler case study. A model-free direct search method and model-based methods using a fundamental model and an empirical efficiency curve model are implemented in the boiler network optimization. The tradeoff between model complexity and plant experimentation is investigated in this case study. The RTO system using a fundamental model can track this changing optimum closely without plant experimentation. Using an empirical efficiency curve model, experimentation is required to track the changing optimum responding to the disturbances in fuel composition and heat exchanger fouling. There is a loss in boiler efficiency during experimentation which is not desirable for tracking fast disturbances. The model-free direct search method takes a lot of steps to reach the optimum, and a significant loss in boiler efficiency is observed during transient.