On Model Identification In The Presence Of Non-stationary Disturbance And Its Use In Model Predictive Control

Over the last few decades, model predictive control (MPC) techniques have beenwidely used in process industries. Since MPC is a model based technique, the modelaccuracy is a key for its successful applications. The system modeling in the presence ofunmeasured disturbances is a complex task. Their presence could affect accuracy of thesystem model, which results in performance degradation for use in MPC. This workaddresses the problem of model identification in the presence of unmeasureddisturbances and its use in MPC techniques.Model identification in the presence of stationary disturbances has been welladdressed in the literature, and these identification techniques are widely used. Most ofthe MPC techniques are based on such models. However, in the process industries, thedisturbances are generally non-stationary which are difficult to model and predict. Thisthesis proposes the three stage identification algorithms for the model identification inthe presence of non-stationary disturbances.The main contributions can be described as follows.An extended Box-Jenkins (EBJ) model is presented, in which a time varying biasterm for representing the non-stationary disturbance is added. This model is capable ofdescribing the dynamics of the process model, and of both the stationary andnon-stationary disturbances. Extension of the proposed model to the MIMO system isalso presented.For the identification of the proposed EBJ model, a three stage multi-innovationrecursive least squares (TSMIRLS) identification algorithm is presented. EBJ model isdecomposed into three subsystems, i.e., the process model, stationary disturbance modeland non-stationary disturbance. Each subsystem is separately identified using amulti-innovation RLS scheme. Each subsystem can be estimated with different tuningparameters. The simulation results show improved accuracy in the parameter estimationfor EBJ model, as compared with BJ model.The parameters of the proposed EBJ model are also identified using a three stagemulti-innovation stochastic gradient (TSMISG) algorithm. As RLS based algorithmshave large computational time, they may not be feasible for estimating systems with alarge number of parameters. In order to reduce the computational burden, the SG approach is used in the three stage identification algorithm. The results for the algorithmare compared and analyzed in terms of computational complexity and accuracy.Based on the above three stage model identification techniques, two different MPCschemes are proposed. Firstly, for the time-invariant systems, an adaptive disturbancemodel is proposed. In this scheme, the process model remains fixed while theparameters of the disturbance model are estimated online. The residual generated as aresult of the output error is used for the estimation of unmeasured disturbances. Fromthe adaptive disturbance model, future disturbances are predicted and included into themodel predictions to improve accuracy. From the simulation example of a glasshouseprocess, the efficiency of the proposed scheme is illustrated. Secondly, an adaptiveMPC for time varying systems is presented where both the process and disturbancemodels are updated online. EBJ model is extended to represent time varying systems. Inthe MPC formulations, a time varying objective function is used. The efficiency of themodel identification technique and adaptive MPC scheme are illustrated through asimulation example of quadruple tank.