This paper addresses the optimization problem of the overall design stage of Hydrogen-powered Unmanned Aerial Vehicles （H-UAVs） in the context of heterogeneous multi-source domains. It explores how to effectively utilize the transfer learning technology to establish a surrogate model and optimize H-UAVs in the presence of heterogeneous samples. To solve the problem of high cost of building a surrogate model due to heterogeneous samples during the evolution of hydrogen-powered UAVs， a framework for establishing a Multi-Source domain Fusion （DG-MSF） surrogate model is proposed based on Data Generation. The geodesic flow kernel method is used to map the heterogeneous source and target domains to a high-dimensional space to determine the relationship between multi-source domains. The marginal distribution-based data generation method is used to effectively integrate source domain information. A multi-layer perceptron neural network is built as a surrogate model， and is trained and fine-tuned through pre-training and fine-tuning methods to achieve efficient prediction of performance of H-UAVs. Finally， the optimization design of the H-UAV is carried out. The analysis results show that the proposed method can effectively utilize multi-source domain data to improve the efficiency of model training and prediction accuracy and the overall performance of H-UAVs， providing powerful technical support for the development of H-UAVs.