Data-driven Method For Model Predictive Controller Design And Performance Monitoring

Model predictive control (MPC) has three key points: namely predictivemodel, rolling optimization and feedback correction. MPC is one of the mostuseful advanced technologies. It can deal with constrains for multivariablesystem so it has been successfully applied in industrial processes. Thereforethe research on design and monitoring for MPC has important theoretical andpractical significance to assure the stable and effective operation.The concept of data-driven originated in computer field and be appliedto control field in recent years. Today, a large number of process data hasaccumulated, how to take full use of the huge information which contained inthe data, to solve the modeling problem in industrial process and assure theeffectiveness of controller, is the key issue for research on data-drivenmethods. It represents a pivotal question in da ta-driven MPC design andmonitoring that an effective and reliable controller which based on a largenumber of input and output process data and a performance monitoring indexwhich can characterize “How healthy” of the current MPC. Centering on these problems, this paper focuses on data-driven MPC design methods andMPC performance monitoring methods, hope to propose some feasiblemethods.Firstly, this paper introduces the research background and meanings.Then reviews the present research status for data-driven control anddata-driven performance monitoring. Based on this, extensive researches areconducted on three parts: MPC design, MPC performance assessment andMPC performance diagnosis.For the data-driven MPC design part, the framework of subspaceidentification algorithm and two common MPC design methods based onsubspace identification are introduced. The NIAT experiment platform isconstructed and two case studies of this platform are used to demonstrate theeffectiveness of the proposed data-driven subspace MPC methods. Thesework provided an experimental verification environment for the research onMPC performance assessment and diagnosis below.For the performance assessment part, under the subspace identificationframework, a subspace matrix rank based method is proposed. Due to thesubspace matrix reflects the dynamic characteristics of the system, the ranksof the subspace matrices are taken as the assessment index with itsbenchmark deduced. A simulation based on the Wood-Berry column process is used to demonstrate the effectiveness of the proposed method. After that,considering the MPC assessment method based on historical benchmark, anew search algorithm for historical data set is proposed, which can overcomethe limitations of previous methods rely on the knowledge of experts, cansearch to get the optimal historical data set based on the needs of differentusers and can improve the sensitivity and accuracy of the assessment index.The simulation results on Wood-Berry column verified the effectiveness ofthe proposed algorithm.For the performance diagnosis part, a performance diagnosis methodwhich based on neural network is proposed. Analysis four degradationfactors for MPC: model mismatch, noise mismatch, input variables andoutput variables constraint saturation. Pattern signatures are extracted fromthe process input and output variables, and classifier is constructed vianeural network. The experiment results on NIAT platform by a two tankliquid level process demonstrate the effectiveness of the proposed method.Subsequently, considering that the performance of MPC may be caused bymultiple factors in the actual production process, the neural networkalgorithm has the advantages in using less priori knowledge and has theability to do f ault diagnosis of complex multi-mode. The proposed neuralnetwork based MPC performance diagnosis method is applied to the case when the MPC performance is caused by multiple factors. A diagnosticsimulation is designed for multiple factors, the simulation results showedthe effectiveness of the proposed method in handling MPC performancediagnostic problem under the influence of multiple factors.