| For years people in the process control field have expressed their interest in having realistic problems for testing process control technology. In response to this interest, several industrial companies have published their test problems. The Tennessee Eastman Control Problem is one of them. The problem provides a plant-wide, realistic, and complicated test process which has large number of manipulated variables and measurements, strong nonlinearities, large disturbances, hard constraints, and open-loop instability. The problem has captured wide attention both in academia and industry. This dissertation investigates a variety of control problems raised by the test problem.In this dissertation, a systematic methodology for control system configuration is developed. The methodology involves screening a large number of candidate control structures using steady state analysis tools. Dynamic simulation is used to achieve a final scheme. With this methodology, a base control system is designed for the Tennessee Eastman process, and it rejects all disturbances, follows all setpoint changes, and satisfies all the process constraints and operating demands.An optimal averaging level control approach, proposed by other researchers, is modified for cases with cyclic disturbances. The modified averaging level control approach is tested and compared with other averaging level control approaches on the Tennessee Eastman process, and the modified averaging level approach, which significantly reduces variation in the product flow, gives the best results. An optimal averaging level control problem for multiple tank systems is also proposed and solved.An inferential parallel cascade control structure is proposed. In the parallel cascade approach, inferential models are used to infer unmeasurable disturbances. The input of the models are secondary measurements, such as the temperatures, pressures, etc., and the outputs of the models are the manipulated variables of primary measurements which are the product properties. For nonlinear processes, neural network partial least squares (NNPLS) models are built. The unmeasurable disturbances affect the secondary measurements faster than they affect the primary measurements, so control of the product properties can be improved. The inferential approach with NNPLS models is applied to the Tennessee Eastman process, and variations in the product flow and compositions are significantly reduced. |
