Research On Nonlinear Iterative Learning Model Predictive Control

Batch processes play an important role in modern intelligent manufacturing industry,which generally undertake the task of manufacturing refined products with high added value.And the product quality highly depends on the tracking accuracy of reference trajectories.Given the repetitive nature of batch operation in finite time period,iterative learning control has become mainstream in the area of batch process control,which updates the control law by learning from the process data of historical batches,so as to progressively improve the tracking accuracy.Iterative learning control is a typical one-dimensional control strategy,with the control input updated only along the iteration axis,and the open-loop control structure adopted in time domain.Therefore,iterative learning control encounters difficulties in dealing with the real-time disturbance and guaranteeing time-domain stability.As an advanced control strategy widely applied in industrial optimization,model predictive control solves the receding horizon optimization problem at each sampling point on account of the predictive states and outputs,thus distinguished by strong anti-interference ability and guaranteed closed-loop stability.Iterative learning model predictive control combines the point-to-point learning mechanism of iterative learning control and the receding horizon optimization framework of model predictive control,establishing a two-dimensional control scheme to realize the targets both in iteration domain and time domain.This organic integration of data-based learning and process control is significant for promoting the process of intellectualization in batch manufacturing,and will be an important step of implementing the "Made in China 2025" policy.The theoretical development of iterative learning model predictive control is still in the early stage.The crucial issues existing in practical batch processes,such as strong nonlinearities,fast dynamics,iteration-varying reference trajectories and trial lengths,bring great challenges to the application of iterative learning model predictive control.In accordance with the multiple production modes of nonlinear batch processes,several iterative learning model predictive control strategies are proposed in this thesis.The stability,robustness and convergence of the presented algorithms are analyzed quantitatively.The main contributions are as follows:(1)A robust iterative learning model predictive control is designed for nonlinear batch process with iteration-varying reference trajectories.The linear parameter varying model is used to approximate the nonlinear dynamics of the controlled system.The H? robust control is embedded in the iterative learning model predictive control to suppress the tracking error fluctuation caused by the variations of reference trajectories.The control input signal is obtained by optimizing the objective function under the constraints of linear matrix inequalities.The stability and convergence of the proposed robust iterative learning model predictive control are theoretically analyzed,while its effectiveness in adapting to changing references is verified by the simulations on a nonlinear numerical example and the stirred tank reactor system.(2)An efficient iterative learning predictive functional control is established for nonlinear batch processes with fast dynamics.The nonlinear system is linearized along the reference trajectory to obtain a two-dimensional predictive model,with the linearized error compensated by formulating the objective function as the upper bound of the actual tracking error.The predictive functional control is applied in time domain to reduce the dimension of the optimization variable,thus effectively relieve the computational burden.In virtue of the terminal constraint set,the time-domain stability and iterative convergence of the proposed iterative learning predictive functional control are proved.The simulations on an unmanned ground vehicle and a fast batch reactor verify that the iterative learning predictive functional control can evidently improve control efficiency meanwhile realize high-precision tracking.(3)An iterative learning model predictive control based on iterative data-driven modeling is proposed for batch processes with complex nonlinearities.The control affine feedforward neural network is established to iteratively identify the nonlinear dynamics utilizing the accumulating process data.Considering the model mismatch derived from neural network modeling errors,an iterative learning model predictive control is designed in tube framework to keep the actual tracking errors within the tube invariant set,thus improving the tracking accuracy.Based on the control affine structure of the neural network predictive model,the gradients of the objective function are computed offline analytically,so that the efficiency and feasibility of online optimization are enhanced.The robust stability and iterative convergence of the data-driven system are proved theoretically.The effectiveness of the data-driven tube iterative learning model predictive control is validated through modeling and control simulations on a nonlinear batch reactor system.(4)An event-triggered iterative learning model predictive control is constructed for nonlinear batch processes with iteration-varying trial lengths.The absent operation information is accurately replenished by the predictive sequences generated by neural network model,which provides each trial with sufficient and reliable historical data.Relying on the triggering condition concerned with the lengths of the adjacent batches,the iterative learning model predictive controller switches between the first-order learning structure and the higher-order learning structure,realizing deep utilization of the real operation information.Under the two control modes divided by the triggering condition,the convergence of the nonlinear iterative learning model predictive control is rigorously proved.The simulations on a numerical system and the injection molding process verify the effectiveness of the proposed method.