<p>Short-term precipitation forecasting plays a crucial role in meteorology, but the complexity of meteorological data and the variability of spatiotemporal features pose challenges to the accuracy of current models. To address issues such as the large cumulative errors and poor extraction of spatiotemporal features in current models, a new Spatio-Temporal Feature Fusion UNet (STFF-UNet) aimed at improving the accuracy of short-term precipitation forecasting is proposed. Using the first three and last three layers of the traditional UNet network as the backbone, this paper introduces a Trans-Time Step Attention Module (TTSAM) and a Cross-Spatial Attention Fusion Module (CSAFM) after the encoding layers. The TTSAM module effectively captures temporal information by focusing on the dynamic changes of important features along the time dimension, while the CSAFM module enhances the model's understanding of complex meteorological patterns by integrating features from different spatial locations through a cross-spatial attention mechanism. A Trans Module (TM) is integrated at the front end of the TTSAM module and the back end of the CSAFM module to further enhance the model's ability to capture and fuse spatiotemporal features. Experiments are carried out using the Storm EVent ImagRy (SEVIR) precipitation forecasting dataset, and the results show that the average CSI of STFF-UNet under different precipitation thresholds is on average 15.7% higher than the other comparison models, which greatly improves the accuracy and reliability of precipitation forecasting. This study indicates that the strategy of spatiotemporal feature fusion can effectively enhance the performance of short-term precipitation forecasting models, providing new insights for future research in precipitation forecasting.</p>