Dynamic parameter estimation of the alkaline electrolysis system combining Bayesian inference and adaptive polynomial surrogate models

Utility-scale hydrogen production via alkaline electrolysis (AEL) is a promising pathway toward the decarbonization of the power, transportation, and chemical industries. The efficiency, load flexibility, and operational safety of the AEL system are subject to electrochemical, thermal, and mass transfer dynamics, and the corresponding parameters, including overvoltage coefficients, heat capacities and resistances of the stack and lye-gas separators, thickness and permeability of the diaphragm, etc. The community has developed many models to depict these dynamic behaviors. However, due to the lack of a comprehensive parameter estimation method, these models are generally tuned manually in industrial applications, which can be inaccurate and cannot fit their time-varying nature. To fill this gap, we present a fast and accurate parameter estimation method for the AEL system. Specifically, to address the difficulties of strong nonlinearity of the dynamic electrolyzer models and correlation between different parameters, a Bayesian inference-based Markov chain Monte Carlo method is proposed. To reduce the computing time for online estimation, data-driven adaptive polynomial surrogate models are established to replace repeated time-domain simulations of the electrolyzer model so that estimation can be finished within a few minutes. Experiments on a 5 Nm3/hr-rated AEL system validate the proposed method. Compared with the existing Kalman filter variants, the estimation error is reduced by at most 71.1% in terms of RMSE and NRMSE. In addition, the proposed method provides approaches to fault diagnosis and global sensitivity analysis for operating and designing AEL systems.