Uncertainty and variability are inherent to chemical processes, driven by factors such as kinetic and thermodynamic correlations, raw material fluctuations, and evolving market demands. With the global shift toward renewable energy, there is an increasing demand for advanced technologies capable of operating flexibly and reliably under dynamic and uncertain conditions. This work focuses on bridging physics-based modeling with data-driven techniques to improve the design and operations of chemical systems, particularly in proton exchange membrane water electrolysis (PEMWE) systems.
This work presents two hybrid modeling strategies. The first is a hybrid mechanistic/data-driven approach developed to predict the performance of PEM electrolyzers and optimize their dynamic operation. This strategy leverages a Physics-Informed Neural Network (PINN) framework, by integrating a physical loss to improve the accuracy of temperature and hydrogen flow rate predictions. The second strategy introduces a hybrid design approach to address feasibility and flexibility in nonlinear process systems. It integrates multi-parametric programming with Y-wise Affine Neural Networks (YANNs) to efficiently characterize feasible operating regions under deterministic uncertainty. This method is scalable to high-dimensional, nonconvex problems and is demonstrated through two case studies, including the flexible design of a PEM electrolyzer. The described approaches advance operability and flexibility modeling by providing a generalized computational framework applicable to linear and nonlinear, convex and nonconvex process systems under deterministic uncertainty.
Keywords: Hybrid Modeling, PEM Water Electrolysis, Flexibility Analysis, Physics-Informed Neural Networks, Parametric Programming