Abstract
Accurate medium-range precipitation forecasting is crucial forhydrometeorological risk management and disaster mitigation, yet remainschallenging for current numerical weather prediction (NWP) systems. Traditionalensemble systems such as the Global Ensemble Forecast System (GEFS) struggle tomaintain high skill, especially for moderate and heavy rainfall at extendedlead times. This study develops a deep learning-based ensemble framework formulti-step precipitation prediction through joint modeling of a comprehensiveset of atmospheric variables. The model is trained on ERA5 reanalysis data at0.25$^{\circ}$ spatial resolution, with precipitation labels from NASA'sIntegrated Multi-satellite Retrievals for Global Precipitation Measurement(GPM) constellation (IMERG), incorporating 57 input variables, includingupper-air and surface predictors. The architecture employs a patch-based SwinTransformer backbone with periodic convolutions to handle longitudinalcontinuity and integrates time and noise embeddings through conditional layernormalization. A dual-branch decoder predicts total precipitation and othervariables, with targeted freezing of encoder-decoder pathways for specializedtraining. Training minimizes a hybrid loss combining the Continuous RankedProbability Score (CRPS) and weighted log1p mean squared error (log1pMSE),balancing probabilistic accuracy and magnitude fidelity. During inference, themodel ingests real-time Global Forecast System (GFS) initial conditions togenerate 15-day forecasts autoregressively. Evaluation against GEFS using IMERGdata demonstrates higher Critical Success Index (CSI) scores at precipitationthresholds of 0.1 mm, 1 mm, 10 mm, and 20 mm, highlighting improved performancefor moderate to heavy rainfall.