Precipitation forecasting holds significant importance in fields such as disaster prevention，mitigation，and water resource management.However，existing precipitation forecasts often encounter issues such as spatial bias and inadequate predictions of extreme precipitation events.This paper proposes a correction model that combines statistical methods with deep learning，named Unet-DConvLSTM-QM （U-D-Q）.The model optimizes precipitation spatial distribution using Unet，employs a dual-layer ConvLSTM to distinguish between rain/no-rain and correct precipitation intensity，and finally applies Quantile Mapping （QM）for further forecast adjustment.Focusing on the Dadu River Basin，the study compares U-D-Q with quantile mapping and ConvLSTM，systematically evaluating its performance.The study's results demonstrate that the U-D-Q model outperforms other methods （quantile mapping and ConvLSTM）in key metrics such as the Critical Success Index （CSI），Mean Absolute Error （MAE），Root Mean Square Error （RMSE），and Correlation Coefficient （CC）.It excels particularly in the medium to high-intensity precipitation intervals （[10，25）， [25，50），[50，+∞）），significantly reducing errors in extreme precipitation events.While the QM model performs well in low-intensity precipitation intervals （[0.1，10）），its adaptability to extreme events is limited.Although ConvLSTM shows some improvements，it does not match the overall effectiveness of the U-D-Q model.This research provides a novel approach to enhancing precipitation forecast accuracy and validates the potential of deep learning applications in precipitation forecasting.