Please use this identifier to cite or link to this item: http://hdl.handle.net/123456789/14487
Title: NONLINEAR MODEL PREDICTIVE CONTROL USING LEARNING MACHINES
Authors: Nisha, M. Germin
Keywords: control algorithms
chemical plant
Model predictive control
Model predictive
Issue Date: Mar-2014
Publisher: Dept. of Electrical Engineering iit Roorkee
Abstract: One among the advanced control algorithms that is receiving a great deal of concern in the process industries, chemical plant and oil refineries is predictive control. Predictive control is a prudent control which retains information about the past process variables and survey the current as well as the upcoming process variables. Internal model control, inferential control and model predictive control are some of the popular predictive controllers. Model predictive control (MPC), which works on the basis of receding horizon control has attracted the process control community due to its capability to handle constraints on process variables, nonlinearities and interactions among process variables, disturbances etc. Model predictive controllers are model- based controllers which rely on dynamic models of the process. Process model plays a key role in model predictive controllers. The more accurate the model the more accurate is the controller. State space models, First Principle models, Hammerstein models, Volterra models etc were used widely to develop accurate dynamics of nonlinear processes which are effort demanding and time consuming. Then artificial neural network (ANN) models turned the attention of MPC users due to their ability to perfectly identify complex nonlinear relationships between dependent and independent variables with less effort. Several researchers have approximated nonlinear models by neural networks besides its lengthy training time, requirement of large training data, poor extrapolation, offset for multistep predictions in the presence of disturbances, over fitting with poor generalization etc. Another widely used machine learning technique introduced by Vapnik is deterministic sparse kernel technique named as support vector machine (SVM). Support vector regression (SVR) models are significant for its accuracy and sparse nature. Another existing machine learning technique introduced by Tipping is probabilistic sparse kernel learning technique called as relevance vector machines (RVM). Relevance vector regression (RVR) models which are much sparser reproduces the nonlinear dynamics accurately. In this thesis, a novel neuro fuzzy technique, Extreme ANFIS is proposed and its significance in achieving accurate model is verified. Thus different machine learning techniques say ANN’s, SVM’s, RVM’s and proposed novel neuro-fuzzy technique are employed to show their suitability to achieve accuracy and computational efficiency. The other concern in model predictive controller is computational cost, as it does prediction and optimization at each sampling instant. This could be overcome by fast prediction and fast optimization techniques. Quasi-Newton methods are known for its fast ii optimization as it skips Hessian matrix computations, but accuracy is less. Particle swarm optimization is an evolutionary algorithm which is meant for its success rate but has long processing time. Processing of conventional particle swarm optimization could be speeded up by particle swarm optimization with controllable random exploration velocity (PSOCREV) technique. This PSO-CREV technique improves the intensity of exploration capability of conventional particle swarm optimization significantly by a time-varying bound of arbitrary search velocity to meet both the necessities of strong exploration skill and fast convergence with less number of iterations and less number of populations. In this thesis, PSO-CREV technique is adopted for its accuracy, simplicity and controllable fast computations. A nonlinear model predictive control strategy which utilizes the above mentioned machine learning techniques and PSO-CREV optimization algorithm is applied to a single input single output (SISO) catalytic continuous stirred tank reactor (CSTR) process. An accurate reliable nonlinear model is first identified by the above mentioned machine learning techniques and then the optimization of control sequence is speeded up by PSOCREV. An improved system performance is guaranteed by an accurate model and an efficient and fast optimization algorithm. Performance comparisons of MPC’s using probabilistic sparse kernel learning technique called RVM’s regression model, deterministic sparse kernel learning technique called Least squares support vector machines (LS-SVM) regression model, a proposed novel neuro-fuzzy based (Extreme adaptive neuro fuzzy inference system (ANFIS)) model and neural network based model are done on a CSTR process. Relevance vector regression model and Extreme ANFIS model shows good tracking performance with very less computation time which is much essential for real time control. A nonlinear system with much faster dynamics is also considered for control. The control of photovoltaic (PV) array Maximum Power Point Tracker (MPPT) through Nonlinear Model Predictive Control (NMPC) strategy which uses Extreme ANFIS/ LSSVM/ RVM regression model is proposed. Another Extreme ANFIS/ LS-SVM/ RVM model is employed to offer the reference Maximum Power Point (MPP) trajectory to the model predictive control system by predicting the maximum power point current and voltage of the nonlinear PV module at different operating conditions. The above control algorithm is speeded up by simplifying the optimization problem by Finite Control Set Model Predictive Control (FCS-MPC) technique. Thus an improved system performance is guaranteed by an accurate predictive model and simple control algorithm. The obtained iii simulation results show the superiority of the proposed method compared to state space model based NMPC. Control of highly nonlinear processes with interacting process variables is still a challenge in industries. Hence, a highly nonlinear binary distillation column process is considered for control to highlight the control accuracy and computational efficiency of NMPC strategy. An accurate reliable nonlinear model is first identified by the proposed novel neuro-fuzzy based (Extreme ANFIS) model and then the optimization of control sequence is speeded up by PSO-CREV. To compare the performance, MPC using probabilistic sparse kernel learning technique RVR with a RBF kernel, deterministic sparse kernel learning technique called LS-SVM regression model and ANN based model is done on a distillation column process. RVR based MPC and Extreme ANFIS based MPC again shows its significance in achieving good tracking performance with very less computational effort which is much essential for real time control applications. Thus this thesis focused in incorporating accurate nonlinear model and reducing the computational cost related to nonlinear model predictive controller.
URI: http://hdl.handle.net/123456789/14487
Appears in Collections:DOCTORAL THESES (Electrical Engg)

Files in This Item:
File Description SizeFormat 
G24344-M. Germin -T.pdf2.9 MBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.