In response to the strong coupling and nonlinear interactions among complex meteorological and marine variables in offshore wind power generation—and given the implicit, topologically intricate nature of multi-source data—this paper introduces a novel multi-source data fusion model that combines a multi-layer attention mechanism (AM) with a bidirectional gated recurrent unit (BiGRU) network. For the spatio-temporal forecasting of offshore wind power, we embed the AM within a deep BiGRU framework to construct a hierarchical attention architecture that jointly learns spatial and temporal dependencies. This architecture dynamically uncovers latent correlations between wind farm outputs and diverse input features, yielding adaptive importance weights across both dimensions. The empirical validation on an offshore wind farm dataset demonstrates that the proposed model achieves superior predictive accuracy and stability compared with benchmark methods.