Application of neural network control to distillation
MetadataShow full item record
Distillation control is challenging due to its coupled, nonlinear, nonstationary, and slow dynamic behavior. Like distillation columns, most chemical processes are usually nonlinear and nonstationary. This nonlinearity greatly limits the effectiveness of linear controllers, especially when the process is operated away from the nominal operating region. Nonlinear controllers, based on phenomenological models, can be developed. However, it is still a very difficult task in real practice, in terms of computational power, to implement these controllers on-line, because the entire model needs to be solved within each control interval. Neural networks give us an alternative approach to model a nonlinear process, and a controller based on this model can overcome the issues of on-line computational problems. Besides nonlinearity, many practical control problems possess constraints on the input, state, and output variables. The ability to handle constraints is essential for any algorithm to be implemented on real processes. Thus strategies for constraint handling within model-based controllers have become one of the more popular research topics. In this dissertation, a constrained optimization technique for control which uses a neural network gain prediction approach has been developed and implemented on a laboratory distillation column as well as on a dynamic simulator. Here, the neural networks are trained based on a phenomenological model. Also, experimental results have been obtained to confirm the applicability of a neural network model-based controller using an inverse of a state-prediction approach that was developed and simulated earlier by Ramchandran and Rhinehart (1994). In addition, two separate single-input-singleoutput (SISO) controllers using the inverse of the state-prediction approach are implemented on the feed and reflux preheaters of the column.