Nonlinear Dynamic System Identification and Model Predictive Control Using Genetic Programming

During the last century, a lot of developments have been made in research of complex nonlinear process control. As a powerful control methodology, model predictive control (MPC) has been extensively applied to chemical industrial applications. Core to MPC is a predictive model of the dynamics of the system being controlled. Most practical systems exhibit complex nonlinear dynamics, which imposes big challenges in system modelling. Being able to automatically evolve both model structure and numeric parameters, Genetic Programming (GP) shows great potential in identifying nonlinear dynamic systems. This thesis is devoted to GP based system identification and model-based control of nonlinear systems. To improve the generalization ability of GP models, a series of experiments that use semantic-based local search within a multiobjective GP framework are reported. The influence of various ways of selecting target subtrees for local search as well as different methods for performing that search were investigated; a comparison with the Random Desired Operator (RDO) of Pawlak et al. was made by statistical hypothesis testing. Compared with the corresponding baseline GP algorithms, models produced by a standard steady state or generational GP followed by a carefully-designed single-objective GP implementing semantic-based local search are statistically more accurate and with smaller (or equal) tree size, compared with the RDO-based GP algorithms. Considering the practical application, how to correctly and efficiently apply an evolved GP model to other larger systems is a critical research concern. Currently, the replication of GP models is normally done by repeating other’s work given the necessary algorithm parameters. However, due to the empirical and stochastic nature of GP, it is difficult to completely reproduce research findings. An XML-based standard file format, named Genetic Programming Markup Language (GPML), is proposed for the interchange of GP trees. A formal definition of this standard and details of implementation are described. GPML provides convenience and modularity for further applications based on GP models. The large-scale adoption of MPC in buildings is not economically viable due to the time and cost involved in designing and adjusting predictive models by expert control engineers. A GP-based control framework is proposed for automatically evolving dynamic nonlinear models for the MPC of buildings. An open-loop system identification was conducted using the data generated by a building simulator, and the obtained GP model was then employed to construct the predictive model for the MPC. The experimental result shows GP is able to produce models that allow the MPC of building to achieve the desired temperature band in a single zone space