Abstract: To address the day-ahead wind power prediction for multiple wind farms and multiple time steps, a deeply space-time-fused short-term power prediction model for multiple wind farms considering the power's temporal evolution of each wind farm and the spatial correlation between the multiple wind farms is proposed. This model consists of gated recurrent units (GRU), multi-core convolutional layers, and a time-varying pattern attention mechanism. First, the temporal and multi-period features of the historical power data in each wind farm are extracted through the GRUs and the multi-core convolution layers, respectively. Then the time-varying pattern attention mechanism is introduced to give correlation weights to the time-varying feature evolution patterns in the multiple wind farms, simultaneously realizing the vertical tracking and the horizontal comparison of the power temporal evolution rules for multiple wind farms. The actual case results of a wind power base in northern China show that the prediction model is able to effectively utilize the wind power spatial and temporal features. It has higher prediction accuracy and stronger learning ability to the wind power time-varying patterns compared with several other popular prediction models.