Abstract
Accidental leakage during offshore hydrogen-blended natural gas transportation
can trigger massive bubble plume migration, though the underlying dispersion
mechanisms and safety risk assessment remain unclear. This thesis aims to
systematically investigate hydrogen blending’s impact on bubble plume dispersion
dynamics and develop plume trajectory prediction models. Motivated by this aim,
small-scale experimental studies and actual-scale numerical investigations, alongside
data-driven machine learning modelling, are implemented.
Effects of leakage size, leakage rate, blending ratios, and leakage directions are
experimentally investigated. Typical bubble plume dynamic structure is analysed, with
two surface flow modes proposed. Higher leakage rates promote a dispersed-tocontinuous bubble plume transition. Increasing hydrogen blending ratio delays this
transition and avoids periodic ruptures. With no blending, plume rupture occurrence
probability drops to 0 at 0.30 MPa, shifting to 0.40 MPa with 20 % blending. Rupture
probability remains 46 % even at 0.50 MPa with 100 % blending. Correlations between
dimensionless penetration length, dimensionless plume offset, and Froude number are
proposed, respectively. Effects of different leakage conditions on related plume
parameters are also discussed. The predictive model of dimensionless maximum
fountain height exhibits strong generalisation with R2 = 0.81 of unknown experimental
data.
According to actual-scale numerical results, low hydrogen blending ratios exert
minimal influence on bubble plume dispersion and potential safety risks. Ocean
currents can cause significant plume offset, thereby substantially prolonging the rising
time. An increased leakage size (i.e., mass leakage rate) can diminish such influences.
Ocean current-induced plume offset reduces the maximum fountain height significantly,
while wind and wave disturbances alter the surface flow contours.
A data-driven machine learning model is developed to predict small-scale
horizontal plume trajectories and real-scale plume trajectories under marine
environmental effects. Catboost demonstrated superior predictive performance and
generalisation capability, achieving R² = 0.96 and 0.81 on their respective independent,
unknown validation data.
These findings advance the understanding of hazard evolution mechanisms in
undersea hydrogen-blended natural gas leakage scenarios, facilitate improved risk
assessment models, and justify applying machine learning methods in this field.
Keywords: Hydrogen safety; Blume plume; Hydrogen blending; Numerical simulation;
Machine learning