为提高降水预报的精度并解决传统数值模式在预报精度和运算效率上的瓶颈问题，通过结合气象大模型和深度学习后处理方法，对陕西省2008—2018年的降水预报进行实例研究。以FourCastNet气象大模型输出的气象要素场为基础，利用贝叶斯优化的卷积神经网络（convolutional neural network, CNN）／长短期记忆网络（long short-term memory, LSTM）算法构建气象要素场-区域降水信息的预训练模型，生成高分辨率的日降水预报。结果表明：该方法在空间分辨率和预报精度上优于传统数值天气预报模式；区域预训练调优预报结果更准确地反映了区域降水的时空分布；基于贝叶斯优化的深度学习后处理算法能够有效缓解初始场偏差对预报的影响。可见，气象大模型结合深度学习后处理算法在降水精准预报中具有显著的应用潜力，为防灾减灾、农业生产及水资源管理提供了科学支持。

To improve the accuracy of precipitation forecasts and address the limitations of traditional numerical weather prediction models in forecast precision and computational efficiency, a meteorological large model was combined with a deep learning post-processing approach was combined. A case study was conducted for precipitation forecasts over Shaanxi Province during 2008—2018. Based on meteorological variable fields output by the FourCastNet model, a pre-trained model mapping meteorological fields to regional precipitation was constructed using Bayesian-optimized convolutional neural networks (CNN)/long short-term memory (LSTM) networks. The results indicate that this method outperforms traditional numerical weather prediction models in terms of spatial resolution and forecast accuracy. The regionally fine-tuned forecasts more accurately capture the spatiotemporal distribution of precipitation. Furthermore, the Bayesian-optimized deep learning post-processing algorithm effectively mitigates the impact of initial field biases on forecast results. These findings demonstrate the significant potential of integrating meteorological large models with deep learning post-processing algorithms for accurate precipitation forecasting, providing scientific support for disaster prevention, agricultural production, and water resource management.