The design of neural networks (NNs) is presented for treating large, linear model predictive control (MPC) applications that are out of reach with available quadratic programming (QP) solvers. First, we introduce a new feedforward network architecture that enables practitioners to obtain offset-free closed-loop performance with NNs. Second, we discuss the data generation procedure to sample the state space relevant to training the NNs based on anticipated online setpoint changes and plant disturbances. Third, we use the input-to-state stability results available in the MPC literature and establish robustness properties of NN controllers. Finally, we present illustrative simulation studies on process control examples. We apply the NN design approach and compare the performance with online QP based MPC on an industrial crude distillation unit model with 252 states, 32 control inputs, and a control-sample horizon length of 140. Parallel computing is used for data generation and graphical processing units are used for network training. Anticipated plant operational scenarios with setpoints and disturbances that may change during operation must be sampled for NN training. After the offline design phase, NNs execute MPC three to five orders of magnitude faster than an available QP solver with less than 1% loss in the closed-loop performance.

提出了用于处理大型线性模型预测控制（MPC）应用的神经网络（NN）设计，而现有的二次编程（QP）求解器无法满足这些应用。首先，我们介绍了一种新的前馈网络体系结构，使从业人员可以使用NN获得无偏移的闭环性能。其次，我们讨论了基于预期的在线设定值变化和工厂干扰，对与训练NN相关的状态空间进行采样的数据生成过程。第三，我们使用MPC文献中提供的输入到状态稳定性结果，并建立NN控制器的鲁棒性。最后，我们介绍了过程控制示例的说明性仿真研究。我们采用NN设计方法，并将其与基于在线QP的MPC在具有252个状态，32个控制输入和140个控制样本水平长度的工业原油蒸馏装置模型上的性能进行比较。并行计算用于数据生成和图形处理单元用于网络训练。必须对具有设定值和运行过程中可能发生变化的干扰的工厂预期运行场景进行采样，以进行NN培训。在离线设计阶段之后，神经网络执行MPC的速度比可用的QP解算器快三到五个数量级，闭环性能损失不到1％。必须对具有设定值和运行过程中可能发生变化的干扰的工厂预期运行场景进行采样，以进行NN培训。在离线设计阶段之后，神经网络执行MPC的速度比可用的QP解算器快三到五个数量级，闭环性能损失不到1％。必须对具有设定值和运行过程中可能发生变化的干扰的工厂预期运行场景进行采样，以进行NN培训。在离线设计阶段之后，神经网络执行MPC的速度比可用的QP解算器快三到五个数量级，闭环性能损失不到1％。