Data-driven methods have been in sight in ocean engineering for short-term vessel speed prediction. However, due to the requirements of large amounts of training data and the inevitable distributed inconsistency among datasets, they are not easy to generalize or achieve superior performance. Specifically, inadequate recorded data of newly in-serviced vessels and ever-changing met-ocean climates making it hard to fulfill an accurate prediction in practice, and it is tough to be improved by merely raising new data-driven models. Taking the case study of Valemax vessels, we uncover that the input–output features have significant differences between datasets, which impair the effectiveness of data-driven models in practical use. With the purpose of designing domain-invariant features for data-driven models, we propose a domain-adapted feature transfer (DAFT), a generalized strategy to narrow the differences of distribution between datasets, which can be combined with a variety of models to form a framework for short-term vessel speed prediction. The superiority and generalization of DAFT are demonstrated through detailed experiments with quantitative and qualitative analyses. In addition, its robustness when faced with insufficient data and the feasibility of using data from sister-ships are also clearly verified. Code is available at https://github.com/Ldiper/DAFT.