Performance Assessment And Diagnosis Of Model Predictive Controller

As the representative of the advanced control, Model Predictive Control (MPC) has beenwidely applied in complex industrial process, performance assessment and diagnosistechnology has great effect to improve the product quality of process and increase productionefficiency. This paper analyzes the performance monitoring problems of MPC at present, onthis basis, three multi-variable model predictive controller performance assessment methodsand a kind of performance diagnosis method based on the similarity of data sets are studied.Considering various constraints in the practical industry process, the performanceassessment of MPC with constraints is presented based on the design principle of themodel-based predictive minimum variance controller. The control increments are used topredict the optimal predictive outputs of the minimum variance controller. The objectiveoptimization function employs quadratic form of optimal predicted outputs and controlincrements weighted, and then it obtains the optimal control law through solving the quadraticprogramming (QP) problem. This performance assessment index not only takes the varianceof control outputs into consideration but also covers the factor of the variance of controlinputs and the control increments, which made it reflect the performance of the modelpredictive control system more authentically. The simulation example on the Wood-Berrybinary distillation column illustrated the validity of the method.An integration element is usually contained in the real process of the controller so as toensure the process zero steady-state error. However, the traditional LQG (Linear QuadraticGaussian) controller uses quadratic objective function weighted by output predictive error andinput and thus lacks the integration element. According to the problem and the analysis of theMPC objective function, this article changes the LQG control objective function to thequadratic form that weighted by output predictive error and control increment, and a LQGbenchmark computing method based on the subspace is presented and used in MPCperformance analysis. Finally, several simulation examples demonstrate the feasibility andeffectiveness of this method. To monitor MPC performance real-timely using process data, a covariance predictionerror based performance monitoring method is proposed. On the basis of analyzing MPCoptimal objective and control structure, a monitored variable set composed of predictionerrors, manipulated variables and process output variables is developed. Then, a covariancebased real-time performance assessment index is presented using a moving window. For theproblem that covariance index has no control limits, a time sequence model for real-timecovariance index is developed. The prediction error of covariance index is monitored to detectMPC performance deterioration. The source of performance deterioration can be located byusing a performance diagnosis method based on data set similarity. Simulations on theWood-Berry binary distillation column demonstrate the effectiveness of the proposed method.