【目的】精确的暴雨临近预测在防灾减灾、工农业生产和交通运输等多方面起着重要作用，对于保障社会经济发展与人民财产安全具有十分重要的意义。然而现有暴雨智能预测方法没有充分考虑暴雨过程、观测以及建模等环节的不确定性问题，阻碍了预测准确性和稳定性的提升。【方法】本文提出基于卡尔曼“估计-校正”循环网络的暴雨临近预测方法，在个别变化理论约束下估计气象状态，并在卡尔曼滤波的指导下依据估计误差进行气象状态校正，以实现准确和可靠的暴雨预测。所提“估计-校正”网络包括个别变化约束的深度状态估计和估计误差指导的气象状态校正2个核心单元；前者根据历史气象状态估计下一时间步的气象状态及误差；后者根据估计误差和观测误差进行气象状态的校正；二者共同提升暴雨预测精度和稳定性。【结果】在ERA5和NCEP数据集上的实验证明，所提方法的暴雨预测准确性指标CSI比所对比的基线方法提升了5%，并以稳定性指标SPREAD≈0.5的成绩取得了良好稳定性。【结论】验证了在深度学习中融合滤波理论缓解不确定性问题的可行性。

[Objectives] Accurate rainstorm prediction plays an important role in disaster prevention and mitigation, industrial and agricultural production, and transportation, making it crucial for safeguarding social and economic development as well as people's property. However, existing intelligent rainstorm prediction methods fail to fully account for the uncertainties inherent in the rainstorm process itself, as well as in observation and modelling, which limits the improvement of prediction accuracy and stability. [Methods] To address this issue, an "estimation-correction" recurrent network based on filtering theory is proposed. This network estimates the meteorological state with the constraints of substantial derivative and corrects the state according to estimation errors, enabling accurate and reliable rainstorm prediction. The "estimation-correction" network consists of two main units, the state estimation unit and the state correction unit. Constrained by substantial derivative, the state estimation unit estimates the meteorological state and error for the next time step based on historical meteorological states. Guided by estimation error, the state correction unit corrects meteorological state by fusing estimation and observation errors. The two units work together to enhance prediction accuracy and stability. [Results] Experiments conducted on the ERA5 and NCEP reanalysis datasets demonstrate that the proposed method improves the Critical Success Index (CSI) of rainstorm prediction by 5% compared to other methods. Furthermore, it achieves good stability, as indicated by a stability metric (SPREAD≈0.5). [Conclusions] These results validate the feasibility of integrating filtering theory with deep learning to address uncertainty in rainstorm prediction.