Accurate sea surface wind speed prediction is an important condition to ensure the navigation safety and energy conservation of ocean ships. In terms of the two problems of difficult spatial feature decoding and low multi-step prediction accuracy of sea surface wind speed predicition in the field of ocean navigation, we design an improved multi-step spatio-temporal prediction method. For the problem of multi-step prediction, we use the lead time strategy to achieve that an individual model learns and distinguishs the data at multiple prediction time. We take the sea surface wind direction as an exogenous variable, take the month, date and time as covariant variables, and combine them with the historical wind speed data to expand sample space. For the problem of spatial features, the residual U-shaped convolutional network with encoder-decoder structure is used to extract and decode the multi-level spatial information, and the lead time features are input into the encoder and decoder at the same time, which strengthens the effect of decoding the deep features into the corresponding prediction time. The 12-hour prediction experiments on global major oil transportation routes show that the proposed method has lower prediction error than the other six prediction methods.