Research On Performance Monitoring And Model-plant Mismatch Detection Of Multivariable MPC Controller

Model Predictive Control(MPC)is the representative of the advanced control technology.With the wide application of MPC in the industry,the requirement of higher performance is more and more important for enterprises.In recent years,performance monitoring technology is a new popular research fields in MPC.In practice,industrial systems are mostly complex multivariable processes.In view of performance monitoring and detection of model-plant mismatch of multivariable MPC controllers,the main work of this paper is as follows:Current model quality index(MQI)can assess model-plant mismatch(MPM).Considering that the performance of MPC is affected by many factors in practice,two indices are applied to realize real-time monitoring of system performance this paper based on MQI: historical performance index based on MPC objective function and covariance index based on model predictive error.The former assesses the whole performance and the latter reflects the influence of the model mismatch and the disturbance.They respond differently to different factors.Combining the re-identified results of the disturbance innovations,we can get preliminary diagnosis why the system performance decreases.The experiment on the Wood-berry tower demonstrates the effectiveness of this method.Considering that MPM is one of the key factors that result in performance deterioration of MPC in practice,this paper presents a method based on model predictive error to detect MPM.If there is no MPM occurs,the predictive error sequence of each output channel can be regarded as white noise.When MPM is detected,a closed-loop subspace identification method is applied to calculate the order of a minimal realization from the deterministic input to the predictive error.According to the order calculated above,we can judge that MPM is caused by the process model or the disturbance model.The simulation result on the Shell tower verifies the feasibility of this method.When MPM occurs in the process model of multivariable MPC,correlation analysis between the predictive error and the manipulated variable of a channel is affected by other manipulated variables and disturbance,thus unable to locate the MPM accurately.Based on the above problem,partial correlation analysis is used to calculate the correlation between the predictive error and the manipulated variable of each channel,under the premise of removing the effect of other manipulated variables and disturbance.The MPM problem is converted to a distribution problem of partial correlation coefficients in a certain interval.Whether a channel is mismatched is judged by observing the distribution graph.The experiment on the Shell tower demonstrates the effectiveness of this method.