This study develops a multi-model fusion quantitative precipitation forecast technique based on the 3D-Unet deep learning model and forecast products from the ECMWF-IFS global model （abbreviated as EC） and the CMA-SH9 regional model, and introduces the integrated gradient (IG) method to enhance model interpretability. The results indicate that the 3D-Unet can effectively improve the performance of raw precipitation forecasts, and the multi-model fusion deep learning model (DL) significantly outperforms the single-model correction model in terms of the TS score for 5 mm·h-1 precipitation. Real-time operational verification shows that the DL can produce substantially better precipitation forecasts than both the original model and the traditional optimal fusion method (OFM). Compared with that of OFM, the TS scores of precipitation forecasts of the DL model at the 5 mm·h-1 and 10 mm·h-1 thresholds are improved by approximately 70.3% and 26.5%, respectively. For the 24 h accumulated precipitation, the DL forecast achieves higher TS scores than a single model and OFM products for moderate rain, heavy rain and torrential rain categories. In particular, relative to that of the EC, at the longer lead time of 60 h and 72 h, TS scores of the DL for torrential rain forecasts are improved by about 42.7% and 24.3%, respectively. The DL can effectively correct the morphology and location errors of large-scale rain belts in raw forecasts, but it tends to overestimate the spatial extent of local extreme heavy precipitation. Feature importance analysis with the IG algorithm reveals that the DL can distinguish the different roles of the same factors in different models. Factors related to boundary layer processes and local dynamic-thermodynamic structures (e.g., CAPE and near-surface wind) possess higher importance in the CMA-SH9, while factors associated with synoptic-scale circulation background and air mass thermal state (e.g., surface pressure and 2 m temperature) contribute more prominently in the EC.