This deliverable presents a methodological framework for the parametric analysis and out-of-domain generalization of gas network simulations under hydrogen blending conditions. The proposed approach combines high-fidelity steady-state physical gas network modelling with a vector-valued Kernel Ridge Regression (KRR) surrogate model, enabling fast prediction of fluid-dynamic variables and subsequent hydrogen concentration tracking through a dedicated mixing module.
The methodology has been tested on both TSO and DSO networks (Snam and InRete real networks). It has been demonstrated on a real DSO network (Riccione, Italy) and progressively tested under increasing parametric dimensionality, from 4 up to 30 independent input parameters, including inlet pressure set-points, hydrogen blending fractions, pipeline roughness, and distributed gas consumptions.
Results show that the surrogate-based workflow drastically reduces computational time while preserving predictive accuracy, enabling large-scale Monte Carlo simulations and Global Sensitivity Analysis. The global sensitivity analysis allows the assessment of the most impactful input parameters on the nodal hydrogen concentration of each node.
The framework allows the identification of homogeneous gas quality areas, mixing zones, and moving concentration boundaries, providing quantitative support for smart hydrogen blending management and network-wide gas quality control.
