Machine Learning for the Design and Control of Power2X Processes with Application to Methanol Synthesis

我们1)开发了新的数值方法，将传统建模和优化方法的优点与数据驱动的机器学习(ML)的能力相结合，并将其用于2)电力化工过程的新的稳健设计和控制方法。甲醇合成被认为是一个具有挑战性的应用实例。显式考虑了由于供给量强烈变化而产生的非线性动力学。跨学科工作计划反映了三个参与小组在实验实现、概念设计、控制和高效算法方面的互补专业知识。在第一步，开发了混合建模方法。它结合了来自无梯度动力学反应堆的实验数据和可用的物理化学洞察力和有效的ML。我们使用通用微分方程和可微的端到端编程范式，允许对未知或昂贵的模型部分使用深度学习，并扩展了混合整数最优控制(MIOC)和最优实验设计(OED)的方法。在第二步，混合模型用于稳健的对象设计。在第一阶段，重点放在均匀混合的等温反应器和非等温空间分布的反应器上。我们希望提高灵活性和对负载和成分变化的容忍度。为此，考虑了缓冲罐和不同类型的可变进料分布的单级和多级反应器。最优配置和最优额定开环控制曲线将通过MIOC使用上述针对给定特性进料曲线随时间变化而开发的动态混合模型来确定。在建立优化问题的过程中，考虑到实际应用的需要，对被控对象的复杂性进行了限制。在第三步中，除了稳健设计外，还设计了一种鲁棒控制策略来补偿对象模型失配和不可预见的干扰，这与上面考虑的名义情况不同。为此，我们使用重复在线优化(NMPC)，并将最新的概念和技术扩展到混合ML模型的特殊情况。单个无梯度反应堆的建模、设计和控制结果将与实验紧密耦合，允许同时进行高效的数据生成和验证。使用现有的机械参考模型，将在硅胶中对更复杂的固定床反应器进行研究。此外，还计划对可能的第二个资助期进行实验验证。我们计划通过嵌入ML的优化和控制在最优决策方面取得进展。我们的混合ML模型执行物理定律，同时显示出通过符号回归从数据中提取新知识的前景。我们开发离线和在线实验设计方法来处理异类数据，并开发健壮的优化方法，以提高ML应用程序的安全性。

We 1) develop novel numerical methods that combine advantages of traditional modeling and optimization approaches with the power of data-driven machine learning (ML) and use them for 2) a new robust design and control methodology for power2chemicals processes. Methanol synthesis is considered as a challenging application example. Nonlinear dynamics due to strongly varying feed supplies are explicitly taken into account. The interdisciplinary work program reflects the complementary expertise of the three participating groups in experimental realization, conceptual design, control, and efficient algorithms.In a first step, a hybrid modeling methodology is developed. It combines experimental data from a gradientless kinetic reactor with available physico-chemical insight and efficient ML. We use universal differential equations and the differentiable end-to-end programming paradigm which allow to use deep learning for unknown or expensive model parts and to extend methods of mixed-integer optimal control (MIOC) and optimal experimental design (OED).In a second step, the hybrid models are used for robust plant design. In the first period the focus is on well-mixed isothermal reactors and nonisothermal spatially distributed reactors. We want to increase flexibility and tolerance against load and composition changes. For this, buffer tanks and different types of single stage and multi stage reactors with variable feed distribution are considered. The optimal configuration and optimal nominal open loop control profiles will be determined with MIOC using the dynamic hybrid models developed above for given characteristic feed profiles over time. Plant complexity will be constrained in the formulation of the optimization problem, in view of practical applicability.In addition to a robust design, in a third step, also a robust control strategy is developed to compensate plant model mismatch and unforeseen disturbances, which differ from the nominal case considered above. For this purpose we use repetitive online optimization (NMPC) and extend state-of-the-art concepts and techniques to the special case of hybrid ML models.Modeling, design and control results for a single gradientless reactor will be closely coupled to experiments, allowing for efficient data generation and validation at the same time. An investigation of more complex fixed-bed reactors will be done in silico using available mechanistic reference models. An experimental validation is planned for a possible second funding period, among others.We plan to generate advances in Optimal Decision Making via ML-embedded optimization and control. Our hybrid ML models enforce physical laws and at the same time show promise to extract new knowledge from data via symbolic regression. We develop offline and online experimental design methods to cope with heterogeneous data and develop robust optimization methods that shall increase safety in ML applications.