Research On Model Predictive Control Algorithm Based On Dual Ideology

Model predictive control(MPC)has been widely studied and applied because of its excellent ability to explicitly deal with constraints.Its core idea is to predict the future dynamics of the system through the system process model,and use rolling optimization to solve the control input to optimize the system performance at the current time.Therefore,the accuracy of the system model will directly affect the control effect.However,the modeling process in the actual system is complex,especially when the system has process noise,measurement noise,or it is difficult to obtain the actual structure and accurate parameters,which will lead to the uncertainty of the prediction model,which generally needs to be reduced by system identification.But there is a conflict between identification and control activities.Dual control idea can maintain a relative balance between system identification activities and control activities.While controlling the system,it can learn the system,so as to obtain the control input that can reduce the future uncertainty of the system.Based on this,for stochastic systems with noise uncertainty,parameter uncertainty and model structure uncertainty,the idea of dual control is introduced into predictive control.In the process of control,the uncertain information is fully used to drive the control target closer to the desired direction and reduce the impact of uncertainty on the system.The main work is as follows:1.Aiming at random systems with unknown noise,a multi-model predictive control algorithm with dual learning characteristics is proposed.Assuming that the structural characteristics of the system noise are known and its variance is in a finite set,using the posterior probability in Bayesian theory,the true value of the noise variance is learned from a set of models while optimizing the control objective.Compared with the traditional model predictive control algorithm,in a system with noise uncertainty,the dual model predictive control algorithm can effectively learn unknown noise parameters while driving the control system to run in the desired direction.2.Aiming at random systems with unknown and time-varying parameters,a model predictive control strategy with learning characteristics that uses system information to reduce future uncertainty is proposed.The strategy uses the estimated parameters obtained by the estimator to parameterize the uncertainty in the system caused by the parameter estimation error,derive a learning control component that can reduce future uncertainty,and add the control component with learning characteristics to the effect in the control input of the system,the control input learning characteristics are given,so that the system optimizes the control target while reducing the uncertainty of the system in the future,and obtains better system performance.3.Aiming at the stochastic nonlinear system with structural model uncertainty,a model predictive control strategy is proposed that reduces the model selection error due to the active learning feature.Some nonlinear systems may have model jumps during operation,such as failures.MPC with active learning features uses current system information to select the model that best describes the evolution of the current system from a set of possible models,and can quickly identify and switch when the model changes.The proposed control strategy approximates and constrains the Bayesian risk of model selection error with Bhattacharyya distance,which increases the rate of correct model selection in a random system with model structure uncertainty.4.Apply the multi-model predictive control algorithm with dual learning characteristics to the vehicle semi-active suspension system to further verify its effectiveness and applicability.