Accurate station‑level precipitation nowcasting is essential for agriculture, water‑resource management, and disaster mitigation, and is typically framed as a time‑series forecasting task. Conventional time‑series methods encounter two major obstacles:
-
Lack of physics‑guided modeling – meteorological variables are treated as a homogeneous set, ignoring their distinct roles in precipitation formation, which leads to predictions that deviate from underlying physical processes.
-
Severe zero inflation – dry intervals dominate the dataset, masking meaningful precipitation patterns and complicating model training.
To tackle these issues, we propose MZ‑Rain, a moisture‑budget‑guided zero‑inflated sLSTM framework. The model is built around the moisture‑budget equation $$\frac{\partial W}{\partial t}= -\nabla\cdot \mathbf{Q}+E-P$$\nwhich decomposes precipitation formation into four process‑specific pathways:
- Moisture storage ($W$)
- Moisture transport ($\mathbf{Q}$)
- Surface evaporation ($E$)
- Precipitation persistence ($P$) Each pathway is modeled by a dedicated sLSTM branch, capturing its temporal evolution while preserving physical consistency.
To address the zero‑inflated nature of precipitation, MZ‑Rain adopts an adaptive Tweedie modeling strategy. It jointly learns the rainfall mean and an auxiliary precipitation‑occurrence task, enabling a balanced trade‑off between dry‑wet discrimination and quantitative precipitation estimation.
Extensive experiments across diverse geographic and climatic regimes show that MZ‑Rain consistently outperforms strong baselines on metrics such as CSI, FAR, MSE, and MAE. The model especially excels at forecasting heavy‑rain events, confirming the benefit of physics‑driven process modeling.
Review: MZ‑Rain elegantly fuses physical insight with deep sequential modeling, overcoming zero inflation and ensuring physical plausibility, and thus sets a new benchmark for station‑level precipitation nowcasting.