| In order for chemical industries to take full advantage of Model Predictive Control (MPC) technique, more and more efforts are directed towards developing identification methods to obtain accurate models of chemical processes. The appearance of subspace identification method has paved the way for easier multivariable process identification, which is critical for broader application of the MPC technique. However, subspace identification does not provide a full answer for all process identification problems in practice. The main challenge in identifying chemical processes is that most chemical processes are inherently nonlinear, some of them involving periodic characteristics, which make classic linear identification techniques inapplicable. In addition, many processes involve significant deadtime, for which the conventional identification approaches tend to lead poor predictions. To overcome these difficulties, the subspace identification method is tailored in several ways to serve the special needs of chemical process control.; For the identification problem of periodic time-varying systems (PTV), an efficient way to identify a PTV system is developed with the help of “lifting” and the subspace identification techniques. This approach overcomes the computational difficulty involved in directly identifying a PTV state-space system and can be easily combined with control and monitoring techniques to realize real-time control and monitoring of periodic processes. To address the problem with significant deadtime, the conventional subspace identification method is modified to put emphasis on obtaining accurate k-step-ahead predictions. This overcomes the failure of the conventional identification approaches when applied to processes, for which accurate long-range prediction is needed. Finally, a novel recurrent feedback neural network structure is suggested to perform nonlinear system identification under unknown disturbance effects. The inclusion of the feedback error term as an input to the model significantly improves the prediction capability of the model under stochastic conditions. |
