Extension of Real-Time Optimisation Using a MultiModel Approach in Conjunction with Evolutionary Algorithms as to Adapt the Structure of These Models

Process optimisation is the discipline of adjusting the inputs and parameters of a process as to improve the performance whilst meeting certain constraints. This is a sub-branch of process control, which is the discipline of ensuring that the desired operation of a process is achieved, which is a branch of chemical engineering. Process optimisation is used in the design stage of a process, where certain parameters, such as flowrates or reactor volumes are modified as to reach the optimum operating conditions. In order to carry out this optimisation, a model of the process is required. Chemical processes are often very complex and even with the best practices of developing models for these processes they are often not fully satisfactory for the accurate prediction of the true processes.
This is where real-time optimisation (RTO) is implemented. This is a technique which uses in situ measurements of the process, along with the inaccurate model, to modify the inputs to the processes as to drive them towards the true optimum. This optimum is defined as the minimum of the cost function, subject to satisfying constraint functions. These functions depend on the process inputs and the model parameters. There are 3 main methods of RTO: The first being model parameter adaption, where certain parameters of the model can be iteratively adjusted as to better fit the plant, however this technique cannot correctly modify the plant as to reach the true optimum in the presence of a structural plant-model mismatch. The second technique is direct input adaption, where the inputs are modified directly in a feed-back control-inspired fashion. This method does not require an iterative model re-optimisation, rather it uses the model to design the controller. The third method is modifier adaption (MA), which adds a linear term to the cost and constraint function as to meet the gradients and values of the true process. This technique's main difficulty is in estimating the gradients of the plant, however if done well enough, MA will reach, upon convergence, the optimum point of the plant (subject to certain model-adequacy conditions which are much less stringent than model parameter adaption).
For these reasons MA will be taken forward and extended as part of my proposed research. Recent advances in MA have developed frameworks to use transient data of the processes and reoptimize before steady state is achieved, which improves the convergence time to the optimal point. Another advancement is in frameworks to optimise closed-loop problems where only open-loop models are available. I intend to combine these two frameworks into a single framework. Next, I will investigate the use of a multiple model technique that has already been implemented in process design, but has yet to be implemented in the context of RTO. Finally, I will conclude my research by investigating the possibility of structural modifications of the population of models via the implementation of an evolutionary algorithm.