Model predictive control (MPC) is multivariate in nature and able to deal with constraints on the inputs, slew rates, etc. MPC reduces the variance of the control system so that the process can operate at the constraint border. So the profitability of process operations is available. For the present time, MPC can be found in a variety of applications including oil, gas, chemical manufacturing and papermaking industries. But industrial practitioners gradually realized that control systems are initially designed using many uncertainties caused by approximations in process models, estimations of disturbance dynamics, and assumptions about operating conditions. The uncertainties may lead to significant difference between plant performance and design specifications. Even if control systems initially perform well, as time passed, many factors can contribute to their abrupt or gradual performance deterioration, such as the change of feedstock, malfunction of sensors or actuators, deactivation of the catalyst and some unmeasured disturbances. To assure effectiveness of process control and consequently safe and profitable plant operation, control performance assessment (CPA) and monitoring (CPM) are necessary. In view of this problem, this dissertation does some researches on CPA/CPM. The main works of the dissertation are follows:1. The background, principle and process of CPA/CPM are briefly introduced, and the development trend and current situation of performance assessment techniques are summarized and concluded. So the theoretical basis for follow-up chapters is laid.2. The approach of multivariate performance assessment based on minimum variance control (MVC) benchmark is analyzed in detail. And this benchmark is used as the first-level performance benchmark of model predictive control. Using MVC benchmark, it provides the useful information that how much potential there is to improve controller performance further.3. To solve the problem that minimum variance control benchmark is not achievable by model predictive control and the calculation of the interactor matrix is complicated, a method based on multi-step prediction error benchmark for performance assessment of MPC system is proposed in this chapter. According to the MPC objective and single step optimal prediction error, optimal prediction performance over multiple prediction points is derived which is employed as the performance benchmark.. This new benchmark can be acquired from the process data and the order of the interactor matrix. Simulation results demonstrate the effectiveness of the given method.4. A novel method based on the historical benchmark for performance assessment and monitoring of model predictive control system is proposed in this chapter. Firstly, the benchmark is obtained by a period of "golden" operation data. Then a performance measure based on the ratio of the historical benchmark and achieved performance is used for the real-time monitoring. When performance changes are detected, data-based covariance monitoring and generalized eigenvalue diagnosis are used to determine the directions and subspaces with significantly worse or better performance versus the benchmark. Some reasons of the performance change are obtained for control engineers. Finally, simulation results show good performances of this method. |