An Autoregressive Diffusion Model for Coarse-Grained Backmapping

To produce green hydrogen in an industrial scale, electrolysis plants are expanding both in quantity and production capacity. Meeting the requirement of renewable power sources within that process, the challenge of fluctuating power availability and price arise. The technology of Proton-Exchange-Membrane (PEM) electrolysis offers a potential solution by enabling fast dynamic load changes, contributing to grid stability while using economically attractive excess power. Alternatively, adapting to a varying product demand defined by the downstream utilization reduces the necessity for hydrogen to be stored in liquid form or pressure tanks via power-consuming process steps. For optimal cost-efficient plant operation, the fluctuating power needs to be balanced against a constant or variable hydrogen consumption.

By using a digital twin to augment plant operation, an increased observability of plant state, equipment performance and integrity is achievable. Hence, critical states of non-measurable conditions can be detected based on data provided by virtual sensors (calculated sensor values from digital twin). For electrolyzers, such critical conditions can be hydrogen crossover, temperature gradients or the state of degradation. Further, it is possible to analyze the plant operation, investigating factors limiting the load change rate and providing suggestions for efficient transitions between operating points. Another possible application field of the digital twin is the use as Operator Training System.

A digital twin, as presented in this work, is a model for the respective electrolysis plant in which the dynamics are incorporated in a mechanistic and multi-physics way. Submodels for each plant component are connected to constitute the respective plant structure. This modular approach enables an easy adaption to other plant configurations, variable model detail in the submodels and the inclusion of proprietary manufacturer models for specific plant components. Additionally, to simulate the operation dynamics correctly, the control logic of the real plant is replicated in the model.

To finally create the digital twin, the dynamic model is coupled with the plant in real-time. For the interaction with the plant, the standardized interface OPC-UA (Open Platform Communications Unified Architecture) for platform independent data exchange is used. The Advanced Process Control System (APCS) of Linde plants supports the OPC-UA interface, thus allowing flexibility for data transfer on both sides. For the digital twin, it is necessary to run the dynamic model independently of the environment in which it is developed. Therefore, the model is compiled as a Functional Mock-up Unit (FMU) and the simulation is executed using Python. The benefit of this architecture lies in the interchangeability of the dynamic model, enabling the utilization of the concept in different plants.

This work shows the development steps for a digital twin by the example of a PEM electrolysis plant. The basis of the respective component models as well as the assembling of the components to a plant model are presented. Specifically, enhancements to a PEM electrolyzer stack model from literature are discussed. Dynamic operation of the PEM electrolyzer model is verified and interface requirements for communication of digital twin and APCS highlighted. Finally, the link between APCS and digital twin is established for data exchange. With this connection, the digital twin replicates the real plant, which is especially interesting for automatic load change (ALC) and automatic start-up (AST) strategies.