Modeling of flood water level prediction using NNARX / Fazlina Ahmat Ruslan

In the wake of fast soft computing processing, advances in hardware interface provisioning and demand over cost efficiency, a substantial acquiesce over the interests in nonlinear control and optimal process operation can been seen in recent academic and industrial study particularly in multivariable control processes. Among such are the distillation processes which are susceptible to various operation perturbations that would directly influence the desired end product outcome. There are several recorded control approaches that have been successfully employed to meet with the end product quality requirement however the current study put the bulk of its focus on a control approach performed based on the reflux ratio as manipulated variable with the top tray temperature as the controlled variable. It is recognized that there is a need to have an offline representation of the pilot plant to allow for anticipation of experiment results to reduce operation error, hence, the main purpose of this study. The pilot plant used as reference process platform in this study is a binary mixture, batch process bubble cap distillation column that suffers from the time-varying nature of its process. It is established that a valid nonlinear multivariable case study is presented by the process plant via a systematic experiment conducted to justify the adopted nonlinear system identification. The implementation of Nonlinear Auto-Regressive with exogenous input (NARX) technique was then explored in this study to prove its reliability as a comprehensive system identification approach which has been cited across various recorded process identification platforms. The present work further investigates the effectiveness of Orthogonal Least Squares (OLS) and Error Reduction Ratio (ERR) as a model structure selection technique to represent a system that is heavily afflicted with intrinsic noise and poor operating conditions. Prescreening method based on Correlation Coefficient analysis was elaborated to reduce the searching pool and comparison studies are presented to seek out the best process output conformity with the reference process dynamics using various well established model selection criteria. Ultimately, the developed identified model is designed and tested such that it can function extensively as an emulation of the plant in an offline environment. Comparison results have shown that the pre-screening method has an essential role in determining an effective process representation especially in real-time multivariable identification framework where a priori knowledge is not available and would help in resultant model generalization performance as opposed to simply using all available model input variables. Further, after a careful investigation into the OLS algorithm, it was shown that the ERR technique which is an essential part of the algorithm to reach model parsimony, has led the resultant model to select an incorrect model terms albeit some improvement in model selection criteria and validation method adopted in this study. This is made apparent when the resultant model was found not being able to generalize a process that deviates from its training parameters.