Optimal Operation of Integrated Chemical Processes: With Application to the Ammonia Synthesis
MetadataShow full item record
Chemical processes have to be operated at their economic optimum to remain competitive. This is called optimal operation. Optimal operation is frequently incorporated using a real-time optimization approach. In this approach, a model of the process is updated using plant and disturbance measurement and subsequently optimized to calculate the setpoints of the controller. The competition in the bulk chemical industry furthermore dictates the necessity of heat and mass integration. As a consequence, it is more and more difficult to obtain a model of the overall process which can be used in real-time optimization. This thesis is therefore applying and developing methods to achieve optimal operation in the case of integrated chemical processes. If it is not possible to obtain a detailed model for real-time optimization of the overall process, it is only natural to try to achieve optimal operation for subprocesses. This can be achieved either through the application of real-time optimization for the respective subprocesses or through process control. The first part of this thesis is investigating different approaches to obtain optimal operation for said subprocesses. Economic nonlinear model predictive control is one approach based on online dynamic optimization. In addition to converging to an optimal operation point, economic nonlinear model predictive control follows the optimal trajectory to this point. Due to the potential complicated optimization problem and problems associated with plant-model mismatch, self-optimizing control in itself and in a hierarchical combination with extremum-seeking control is subsequently applied to the same case study. When disturbances occur, self-optimizing control is keeping the operation close to the optimum whereas extremum-seeking control adjusts the setpoints to the self-optimizing variables to remove the steady-state loss of self-optimizing control. This allows achieving optimal operation without the necessity of a detailed model and reduces the impact of plant-model mismatch and the solution time of the optimization problem in economic nonlinear model predictive control. Feedback real-time optimization as a novel alternative transforms the optimization problem of conventional real-time optimization to a control problem. This allows a fast response to disturbances and removes problems associated with the cost measurement and gradient estimation in extremum-seeking control. The second part of this thesis is developing methods for surrogate model generation. Surrogate models are computational cheap regression models of computational expensive detailed models. They can be used in the context of real-time optimization to reduce the computational load to solve the optimization problem. The separation of the initial process model into subprocesses results in models that are computational cheaper to solve. Surrogate models are subsequently fitted to the subprocesses and combined into a surrogate model flowsheet. The optimization is then performed using the surrogate model flowsheet. Using surrogate models, it is possible to perform a variable transformation. Partial least squares regression allows a reduction in the number of independent variables through the calculation of new, latent variables. The application of self-optimizing variables in the generation of surrogate models results in a simplified surface of the detailed model. The simpler surface requires then fewer points to achieve a satisfactory fit of the surrogate model. Partial least squares regression can be used as a termination criterion in sampling without the need of fitting a surrogate model at each sampling iteration step as well. This results in a reduction of the computational load in sampling for surrogate model generation. The surrogate model flowsheet can then be used in real-time optimization.