The performance of model predictive controllers (MPCs) strongly depends on the precision of the prediction model. Nonlinear systems, such as neutralization reactors, provide special challenges to MPC design. Linear prediction models may be inadequate to describe the process at all operating points. One alternative is the use of artificial neural networks (ANNs) as prediction models. ANNs are nonlinear structures that can be trained to reproduce the process behavior. Inside MPC schemes, ANNs can rapidly predict the process response to a control action. The time-consuming step for ANN training is to obtain a representative overall data set from experiments or simulation data from the studied process. In the present work, we propose to obtain this data set from computational simulations using a first principles model. However, mismatches were found between rigorous simulation and actual pH process responses. Those deviations were naturally transferred to the internal neural model, as a consequence, actual control problems were identified. Avoiding high costs of performing actual experimental runs for ANN and MPC design, we used a real-time adaptation algorithm, based on extended Kalman filter (EKF), that acts to correct the ANN prediction while process is running. The adaptive model ANN-based MPC was able to maintain the actual controlled process, in all operating conditions tested. The sum of square error of pH was reduced in 64.3%, compared to the ANN-based MPC without model adaptation. Using a Kalman filter to adapt the internal model has significantly improved the MPC performance, reducing oscillations and maintaining the controlled variable in the setpoint, even in servo regulatory situations of industrial practice. In addition, the proposed scheme has great potential for controlling highly nonlinear processes.

模型预测控制器（MPC）的性能在很大程度上取决于预测模型的精度。非线性系统，例如中和反应器，对MPC设计提出了特殊的挑战。线性预测模型可能不足以描述所有操作点上的过程。一种替代方法是使用人工神经网络（ANN）作为预测模型。人工神经网络是可以训练以重现过程行为的非线性结构。在MPC方案内部，人工神经网络可以快速预测过程对控制动作的响应。ANN训练的耗时步骤是从实验中获得具有代表性的总体数据集，或从研究过程中获得仿真数据。在当前的工作中，我们建议使用第一原理模型从计算仿真中获取此数据集。然而，在严格的模拟与实际pH过程响应之间发现不匹配。这些偏差自然会转移到内部神经模型中，因此，确定了实际的控制问题。为避免为ANN和MPC设计执行实际实验运行而付出的高昂成本，我们使用了基于扩展卡尔曼滤波器（EKF）的实时自适应算法，该算法可在过程运行时纠正ANN预测。基于ANN的自适应模型MPC能够在所有测试的运行条件下维持实际的受控过程。与没有模型调整的基于ANN的MPC相比，pH的平方误差总和降低了64.3％。使用卡尔曼滤波器调整内部模型可以显着改善MPC性能，减少振荡并将控制变量保持在设定点，即使在工业实践中的伺服调节情况下也是如此。另外，所提出的方案在控制高度非线性过程中具有很大的潜力。