The cost of simplification: Comparing surrogate electrolyser models in a MILP framework

In this work, we investigate how various surrogate models representing an electrolyser’s production curve influence mixed-integer linear programming (MILP) optimisation outcomes. We compare linear, piecewise-linear, and scalable piecewise-linear models, each fitted to a physical electrolyser model, and analyse their effects on hydrogen production overestimation and real revenue in a solar farm/electrolyser array case study. Additionally, we employ a MILP implementation of a rectified linear unit (ReLU) neural network as a surrogate model. The results indicate that using simple linear models can significantly overestimate hydrogen production due to their inability to capture the non-linearity of the component’s efficiency. This overestimation correlates with reduced real revenues. In contrast, the adoption of simple piecewise models can boost returns by up to 14% in this scenario. The findings indicate that simple enhancements to the detail of component models can substantially improve system-level performance and yield more reliable results, without greatly increasing computational demands. While electrolysers were used as a case study, this analysis can be extrapolated and applied to additional components for which surrogate models are required in order to be included in larger energy system models. • The effects of surrogate model choices of an electrolyser component are studied. • Piecewise-linear fits and ReLU neural networks are implemented as surrogate models. • Ignoring electrolyser production non-linearity affects MILP optimisation outcomes. • Linear models can significantly overestimate hydrogen production and reduce revenue. • Study findings are applicable to non-linear components across various energy systems.