Training-only baseline estimation for cross-composition leak characterization in hydrogen-blended natural gas pipelines

Hydrogen blending changes the normal pressure–temperature state of natural gas pipelines, which can confound leak characterization when prediction models are transferred across compositions. This study evaluates whether the no-leak baseline of a held-out hydrogen composition can be estimated using only no-leak references from the compositions available during training. A 188.25 km pipeline was simulated in LedaFlow at 0, 10, and 20 mol% H 2 , generating 135 leak cases from five leak diameters and nine locations, together with three no-leak references. Boundary pressure and temperature signals were uniformly resampled at 10 s and evaluated using leave-one-composition-out regression without using H 2 fraction as a direct regression predictor; the held-out composition was used only as the coordinate for the training-derived baseline estimate. The H00–H02 no-leak shift reached 8.14 bar for inlet pressure and 1.50°C for outlet temperature, exceeding the corresponding median leak-induced responses. Exact composition-matched baseline correction was treated as an oracle reference because it used the held-out no-leak case. In contrast, a linear baseline estimated only from the two training compositions improved all six held-out composition–target evaluations, yielding diameter MAEs of 0.196–0.504 cm and location MAEs of 13.1–21.2 km. ExtraTrees ranked first in all training-only inner cross-validations, while dummy and shuffled-target controls were consistently inferior. Nested target-grid tests showed weaker transfer to previously unseen locations, with mean location MAEs of 20.3 km for interior targets and 37.3 km at the spatial boundaries. Controlled perturbations revealed substantial absolute sensitivity to sensor noise and drift; however, training-only baseline correction retained lower MAE than Raw P/T in paired tests using identical perturbed signals across all tested Gaussian-noise levels and the prescribed linear drift. These results support a relative benefit of training-only baseline estimation within the examined single-system simulation design, while the large absolute error increases under perturbation preclude claims of measurement robustness or field performance.