一种无约束多步递归神经网络预测控制器

Multistep recurrent neural network model predictive controller without constraints

本文针对具有强非线性、多工作点特性的控制系统,提出了一种基于递归BP神经网络的多步预测模型;通过分析预测模型的内在数学关系,选择了二次型函数作为预测控制器的目标函数,并给出了目标函数关于控制序列的雅可比矩阵和赫森矩阵的计算方法;最后使用Newton-Rhapson算法设计出了滚动优化控制策略,构建了一个非线性多步预测控制器.仿真结果表明,文中提出的多步预测控制器具有较好的控制效果.

This paper brings forward a multistep predictive model based on the recurrent backpropagation （BP） neural network for the control systems with strong nonlinearity and multiple set-points. By analyzing the internal mathematical relation of the predictive model, we select a quadratic function as the objective function for the multistep predictive con- troller. For this objective function, we compute the Jacobian matrix and Hessian matrix of the control sequence, and design the receding horizon optimization strategy using Newton-Rhapson algorithm, thus, constituting a nonlinear multistep model oredictive controller. Simulation results show desirable performances of the model predictive controller.