Domain adaptation based high-fidelity prediction for hydrogen-blended natural gas leakage and dispersion

Hydrogen blended natural gas is regarded as an important solution to facilitate the large-scale transmission and utilization of renewable hydrogen energy in the global energy transition. It is particularly susceptible to accidental leakage and dispersion due to the high leakage propensity of both hydrogen and natural gas, which may lead to significant casualties and economic losses. Deep learning approaches have been applied to high-fidelity prediction of accidental leakage and dispersion scenarios, but they exhibit low efficiency and limited generalization for large-scale emerging hydrogen energy scenarios due to the requirements of computationally intensive CFD simulations. This study proposes a domain adaptation based high-fidelity plume prediction model that integrating numerous low-fidelity Gaussian plumes to extract shared plume features, thereby enhancing efficiency and generalization with a limited number of high-fidelity CFD plumes. Numerical simulations for hydrogen blended natural gas leakage and dispersion, including CFD model and Gaussian plume model, are conducted to construct benchmark high and low-fidelity plumes. By using such datasets, the weight combination with shared features weight of λ₂ = 1e-4 and low-fidelity features weight of λ₁ = 1e-4, as well as the number of CFD plumes n = 16 was determined to optimize the proposed model's efficiency and generalization. A comparison between the proposed model and the state-of-the-art models was also conducted. The results demonstrate that the proposed model maintains high prediction accuracy for high-fidelity plumes while reducing CFD computation by 80 %, and surpassing the pre-trained transfer learning model. Overall, the proposed model facilitates large-scale adaptation of deep learning prediction model to various emerging hydrogen energy scenarios, effectively managing the accidental leakage and dispersion risk in renewable hydrogen systems.