Statistical Machine Learning for Model Predictive Control of Nonlinear Processes

近年来，机器学习（ML）因其在大数据集（“大数据”）中发现模式的能力及其在经典工程领域的广泛应用而引起了越来越多的关注。传统上，过程控制系统依赖于线性数据驱动模型，在某些情况下依赖于第一性原理模型。然而，建模大规模，复杂的非线性过程仍然是过程系统工程的主要挑战。由于过程模型是先进的基于模型的控制系统的关键要素，例如模型预测控制（MPC）和经济MPC，因此构建，培训和表征ML模型的准确性是控制系统设计的新前沿，将影响下一代工业控制系统。受此激励，本研究计划的目标是采用并进一步推进机器学习理论中广义误差界限的方法框架，以开发和验证机器学习模型，并将这些模型集成到用于广泛非线性化学过程的预测控制系统设计中。研究成果和软件工具将被纳入加州大学洛杉矶分校的本科过程控制和高级设计/过程经济学课程，以根据部门和校园目标向学生介绍机器学习技术的应用。该项目还将通过参与加州大学洛杉矶分校工程教育和多样性中心，向高中学生和教师，以及波莫纳加州州立理工大学和西班牙裔为主的El-Camino学院的推广，让不同群体的本科生和研究生参与研究。提出的研究计划的目标是采用和推进机器学习理论中广义误差界限的方法框架，用于开发和验证具有特定理论精度保证的机器学习模型，并将这些模型集成到非线性化学过程的模型预测控制（MPC）和经济MPC系统设计中。具体而言，这项研究将侧重于以下广泛目标：A)机器学习模型的广义概率误差界限的发展，考虑神经元和层数对精度的影响，并指导网络结构的选择和训练；b)模型预测控制方案的设计，以过程结构感知递归神经网络的形式纳入机器学习模型，这些模型具有计算效率，并确保所需的闭环稳定性，性能，鲁棒性和操作安全性，c)在模型预测控制中开发机器学习模型的在线自适应方法，以利用实时噪声数据捕获变化的过程动力学和模型不确定性，d)机器学习建模和控制方法的高保真应用，大型过程模拟器，结合了工业合作者提供的数据集的过程数据、参数和噪声，以及研究小组为减少CO和CO2而开发的实验电化学反应器系统。虽然该研究将在化学过程控制系统合成的背景下进行，但最终的设计框架将广泛影响广泛的工业部门的制造过程，以及目前使用模型预测控制的智能设备。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Machine learning (ML) has attracted increased attention in recent years due to its ability to uncover patterns in large sets of data (“big data") and its widening application in classical engineering fields. Traditionally, process control systems rely on a linear data-driven models, and in certain cases on first-principles models. However, modeling large-scale, complex nonlinear processes continues to be a major challenge in process systems engineering. Since process models are key elements of advanced model-based control systems, such as model predictive control (MPC) and economic MPC, building, training, and characterizing the accuracy of ML models is a new frontier in control system design that will impact the next generation of industrial control systems. Motivated by this, the goal of the proposed research program is to employ and further advance the methodological framework of generalized error bounds from machine learning theory for the development and verification of machine learning models and to integrate these models into predictive control system design for broad classes of nonlinear chemical processes. The research results and software tools will be incorporated within the undergraduate process control and senior design/process economics course curricula at UCLA to introduce students to the applications of machine learning techniques in accordance with departmental and campus goals. The project will also involve a diverse group of undergraduate and graduate students in the research through participation in the Center for Engineering Education and Diversity at UCLA, outreach to high school students and teachers, and outreach to the California State Polytechnic University in Pomona and the predominantly Hispanic El-Camino College. The goal of the proposed research program is to employ and advance the methodological framework of generalized error bounds from machine learning theory for the development and verification of machine learning models with specific theoretical accuracy guarantees and integrate these models into model predictive control (MPC) and economic MPC system design for nonlinear chemical processes. Specifically, this research will focus on the following broad objectives: a) the development of generalized probabilistic error bounds for machine learning models accounting for the impact of the number of neurons and layers on accuracy and guiding network structure selection and training, b) the design of model predictive control schemes that incorporate machine learning models in the form of process-structure aware recursive neural networks that are computationally efficient and ensure desired closed-loop stability, performance, robustness and operational safety properties, c) the development of a methodology for on-line adaptation of machine learning models in model predictive control to capture changing process dynamics and model uncertainty using real-time noisy data, and d) applications of the machine learning modeling and control methods to high-fidelity, large-scale process simulators incorporating process-data informed parameters and noise from data sets provided by industrial collaborators as well as from an experimental electrochemical reactor system developed within the research team for the reduction of CO and CO2. While the research will be carried out in the context of chemical process control systems synthesis, the resulting design framework will broadly impact manufacturing processes across a wide range of industrial sectors as well as smart devices that currently make use of model predictive control.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.