This study develops a multi-model fusion quantitative precipitation forecast technique based on the 3D-Unet deep learning model. The method integrates upper-air and surface variables from the ECMWF-IFS global model and the CMA-SH9 regional model for the period 2019–2023. Furthermore, the Integrated Gradients (IG) method is introduced to enhance the interpretability of the model. The results indicate: (1) The 3D-Unet model can effectively enhance the performance of the original precipitation forecast of the model, and the multi-model fusion deep learning forecast model (DL) significantly outperforms the single-model correction model in terms of TS score for 5mm/h precipitation forecast. (2) Real-time operational verification results show that the precipitation forecast performance of the DL model is significantly better than that of the original model forecast and the traditional optimal weight fusion method (OFM). The TS scores for hourly precipitation forecasts of 5mm/h and 10mm/h have increased by approximately 70% and 26.8%, respectively, compared to OFM. For 24-hour cumulative precipitation forecasts, the TS scores of the DL model for moderate rain, heavy rain, and rainstorm levels are superior to those of single models and OFM products; especially for longer forecast horizons of 60 hours and 72 hours, its TS scores for rainstorms have increased by approximately 42.7% and 24.3% compared to the EC model, respectively. (3) The DL model can effectively correct the shape and location of the original forecast rain belt of the model (such as the southwest vortex rainstorm), but it tends to have a wider range of local extreme heavy precipitation. (4) Using the IG algorithm to analyze the importance of feature factors, it is found that the DL model can identify the differences in the effects of the same factors across different models. Factors related to boundary layer processes and local dynamic and thermal structures (such as CAPE and near-surface wind field) have higher importance in the CMA-SH9 regional model, while factors related to synoptic-scale circulation background and air mass thermal state (such as surface pressure and 2m temperature) contribute more prominently in the EC global model.