基于动态反馈神经网络的复杂系统预测控制

Model predictive control for complex systems based on dynamic recurrent neural network

在分析基于动态反馈神经网络(DRNN:DynamicRecurrentNeuralNetwork)的模型预测控制策略的基础上,为改善Elman网络辨识高阶系统时的计算复杂性,采用具有局部动态反馈特性的Elman网络进行线性系统状态空间模型的在线辨识。基于跟踪器型性能指标的预测控制器对系统进行滚动优化,并对动态反馈神经网络逼近状态空间模型进行了证明。对过程控制装置三容系统进行了仿真研究,通过离线训练方式获得网络初值的选择。仿真结果表明,此算法能使系统的输出保持期望轨迹,并能有效处理系统本身的输入、输出约束条件。

DRNN(Dynamic Recurrent Neural Network) based MPC(Model Predictive Control) strategy is discussed. To improve the computational complexity of the Elman network's identification, linear state-space model's online identification based on local dynamic recurrent Elman network is proposed. Optimization strategy is applied with quadratic performance in predictive controller based on quadratic performance, and the approximation of the DNN to DSS(Discrete State Space) model is proved. Algorithm's validity is approved with a three-tank system with the system output can track a desired trajectory and within the limitation.