Scholarlink

<p>At present, state-of-the-art predictions for the Madden-Julian Oscillation (MJO) combine forecasts from numerical weather models together with bias-correction methods. Recently, deep learning (DL) models have outperformed numerical models on global benchmarks, but their potential for improved MJO prediction remains underexplored. Here, we address two key gaps in DL-based MJO prediction: (1) whether incorporating extended historical context improves forecast skill, and (2) whether bias-correction methods are as effective for DL models as for numerical models. Our results show that MJO predictability is governed by the current MJO state rather than prior propagation. Unlike in numerical models, where bias correction significantly improves forecasts across all lead times, bias correction in DL models only takes effect beyond 3 weeks, reducing errors by 11% in weeks 5–6 and extending the skillful horizon from 37 to 40 days. We conclude that existing bias-correction methods are most effective at correcting dataset inconsistencies rather than model deficiencies.</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

On the importance of historical context and bias correction in deep learning MJO prediction models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Khalil Virji,
  • Geruo A,
  • Isabella Velicogna

摘要

At present, state-of-the-art predictions for the Madden-Julian Oscillation (MJO) combine forecasts from numerical weather models together with bias-correction methods. Recently, deep learning (DL) models have outperformed numerical models on global benchmarks, but their potential for improved MJO prediction remains underexplored. Here, we address two key gaps in DL-based MJO prediction: (1) whether incorporating extended historical context improves forecast skill, and (2) whether bias-correction methods are as effective for DL models as for numerical models. Our results show that MJO predictability is governed by the current MJO state rather than prior propagation. Unlike in numerical models, where bias correction significantly improves forecasts across all lead times, bias correction in DL models only takes effect beyond 3 weeks, reducing errors by 11% in weeks 5–6 and extending the skillful horizon from 37 to 40 days. We conclude that existing bias-correction methods are most effective at correcting dataset inconsistencies rather than model deficiencies.