EVENT DOSSIER
On September 7, 2026, the MZ-Rain model paper was published in arXiv cs.AI. This study proposes a zero-expansion sLSTM framework guided by moisture budget, used for short-term forecasting of precipitation at the site level. The model breaks down the precipitation formation process into pathways such as moisture storage, transport, evaporation, and persistence based on the moisture budget equation, and captures temporal evolution through dedicated branches. An adaptive Tweedie modeling strategy is also introduced to adjust the average rainfall value while jointly learning precipitation occurrence as an auxiliary task, thereby balancing dry-wet discrimination and quantitative estimation. Extensive experiments across different geographical and climatic regions show that MZ-Rain outperforms strong baseline models in multiple metrics, especially in forecasting heavy rainfall events.
提出 MZ-Rain 模型，该基于水分预算引导的零膨胀 sLSTM 框架用于站点级降水短时预报。针对传统时间序列建模缺乏物理引导及降水数据严重零膨胀两大挑战，MZ-Rain 依据水分预算方程将降水形成过程分解为水分储存、输送、蒸发及持续等特定路径，并通过专用分支捕捉其时间演化；同时引入自适应 Tweedie 建模策略，在联合学习降水发生作为辅助任务的同时调节降雨均值，以平衡干湿判别与定量估算。跨不同地理和气候区域的广泛实验表明，MZ-Rain 在 CSI、FAR、MSE 及 MAE 等多项指标上均优于强基线模型，尤其在预报强降雨事件方面表现优异。