Scholarlink

<p>Accurate medium-range precipitation forecasting is important for hydrometeorological risk management and disaster mitigation, yet remains challenging for numerical weather prediction and data-driven systems, particularly for precipitation intensity, spatial structure, and ensemble reliability at longer lead times. This study develops CSU-PCAST, a deep learning-based ensemble forecasting framework for global 6-h precipitation prediction. The model is trained with ERA5 atmospheric and surface variables at 0.25° spatial resolution and IMERG precipitation labels, and it uses 57 prognostic atmospheric and surface variables together with static geographical fields. CSU-PCAST predicts both the evolving atmospheric state and 6-h total precipitation through a patch-based Swin Transformer backbone, periodically padded residual convolutions for longitudinal continuity, stochastic noise conditioning, time embeddings, and a dual-branch decoder for precipitation and non-precipitation variables. During inference, CSU-PCAST is initialized from operational GFS analyses and generates 30 autoregressive ensemble members to 15 days, matching the operational GEFS ensemble size. Evaluation over 2023, using IMERG as the precipitation reference, shows improved short-lead deterministic skill relative to GEFS, with higher CSI and lower RMSE during the first several forecast days. CSU-PCAST also reduces GEFS wet bias for light precipitation and dry bias at the 10 and 20 mm thresholds. Probabilistic verification shows lower CRPS, higher Brier Skill Scores for several thresholds, and improved ensemble reliability, although both systems remain underdispersive. A Sanba extreme precipitation case study further demonstrates improved spatial structure and exceedance-probability guidance, while regional precipitation totals, high-end intensity, and ensemble calibration remain areas for further improvement.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CSU-PCAST: a dual-branch transformer framework for medium-range ensemble precipitation forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Tianyi Xiong,
  • Haonan Chen,
  • Kelly Mahoney,
  • Jingyin Tang,
  • Tim Smith,
  • Janice Bytheway

摘要

Accurate medium-range precipitation forecasting is important for hydrometeorological risk management and disaster mitigation, yet remains challenging for numerical weather prediction and data-driven systems, particularly for precipitation intensity, spatial structure, and ensemble reliability at longer lead times. This study develops CSU-PCAST, a deep learning-based ensemble forecasting framework for global 6-h precipitation prediction. The model is trained with ERA5 atmospheric and surface variables at 0.25° spatial resolution and IMERG precipitation labels, and it uses 57 prognostic atmospheric and surface variables together with static geographical fields. CSU-PCAST predicts both the evolving atmospheric state and 6-h total precipitation through a patch-based Swin Transformer backbone, periodically padded residual convolutions for longitudinal continuity, stochastic noise conditioning, time embeddings, and a dual-branch decoder for precipitation and non-precipitation variables. During inference, CSU-PCAST is initialized from operational GFS analyses and generates 30 autoregressive ensemble members to 15 days, matching the operational GEFS ensemble size. Evaluation over 2023, using IMERG as the precipitation reference, shows improved short-lead deterministic skill relative to GEFS, with higher CSI and lower RMSE during the first several forecast days. CSU-PCAST also reduces GEFS wet bias for light precipitation and dry bias at the 10 and 20 mm thresholds. Probabilistic verification shows lower CRPS, higher Brier Skill Scores for several thresholds, and improved ensemble reliability, although both systems remain underdispersive. A Sanba extreme precipitation case study further demonstrates improved spatial structure and exceedance-probability guidance, while regional precipitation totals, high-end intensity, and ensemble calibration remain areas for further improvement.