Traditional numerical weather prediction (NWP) models have inherent biases in quantitative precipitation forecasting tasks due to limited spatial resolution, incomplete physical parameterization schemes, and weak generalization, Deep learning neural networks, with their strong non-linear fitting capabilities, autonomous learning of task-relevant features, and high generalization, hold the potential to address these issues and improve the current state of NWP forecasting. To address the aforementioned issues, this paper proposes a research on short-term quantitative precipitation forecasting technology based on multi factor 3D feature extraction. Based on high-resolution ECMWF HRES (EC Hs) model forecasting data provided by the European Centre for Medium Range Weather Forecasts (ECMWF), a 3D-QPF (3D-Quantitative Precision Forecast) semantic segmentation model is constructed. Through a coupled framework of classification and regression, the 3D spatial features of various precipitation related element data are captured, And obtain a non-linear relationship with the actual precipitation data, and finally add the PR (Pre and Rec) loss function to the loss function to further improve the predictive performance of the model on skewed data.The experimental results show that the accuracy score of 3D-QPF daily cumulative precipitation forecast not only increases steadily on the sunny and rainy scale (0.1mm/24h), but also improves significantly on the rainstorm scale (50mm/24h): the rainstorm scale is 15.8% higher than the EC-Hres 'TS(Threat Score) score, and the root mean square error (RMSE，Root Mean Square Error) optimization reaches 18.71%. After long-term testing, the 3D- QPF model has achieved effective prediction correction compared to classic network models such as ECHres, China Meteorological Administration Global Model (CMA-GFS) forecasting, 2D Unet, and 3D-Unet. In addition, as the forecast time is extended to 3 days, the optimization effect of the model can still remain relatively stable.