采用Brent优化的核学习单步预测控制算法

Kernel learning one-step-ahead predictive control algorithm using Brent optimization

针对非线性SISO系统,提出一种基于核学习辨识模型的单步预测控制算法(kernel learning one-step-ahead predictive control,KLOPC).通过KL辨识模型得到系统的一步超前预报值,并引入输出反馈和偏差校正以克服模型失配等因素引起的预测误差,以此构造一步加权预测控制性能指标,然后采用Brent一维搜索方法求取控制律.该方法无需任何相关的导数信息,需调整的参数少,求解效率高.在一非线性液位系统的仿真研究表明了KLOPC优于整定的PID和其它基于KL模型的控制方法,对噪声和扰动等均具有更好的鲁棒性和自适应性.

A novel kernel learning one-step-ahead predictive control （KLOPC） algorithm is presented for the general unknown single-input/single-output nonlinear systems. Firstly, a one-step-ahead predictive model is obtained by using the KL identification framework; secondly, a new one-step-ahead weighted predictive control performance index is formulated; thirdly, the control law is computed via Brent optimization method, which is efficient and reliable in one dimension search without knowing any derivative of the KL identification model. This simple KLOPC scheme has few parameters to be chosen, making it very suitable for real-time control. Simulation results of a nonlinear process show that the new KLOPC algorithm is superior to other methods based on KL model and the well tuned PID controller. The proposed KLOPC strategy also exhibits more satisfactory robustness and adaptation to both additive noise and unknown process disturbance.