Scholarlink

<p>The risk of heatwaves (HWs) in the late spring–early summer period keeps escalating in Southeast Asia (SEA). To mitigate heatwave (HW) risks effectively, more accurate and longer-term forecasts are urgently needed. This study evaluates two machine learning weather prediction (MLWP) models, the FourCastNet v2 (FCN2) and the Artificial Intelligence Forecasting System (AIFS), in hindcasting the occurrence of HWs in SEA. The SEA HWs are classified into three categories using the K-means clustering: Category 1 for uniform warming across the SEA, Category 2 for pronounced warming over the Maritime Continent (MC), and Category 3 for intensified warming over the mainland SEA. Each category is driven by anomalous descent that promotes air temperature rises in their respective area. The AIFS demonstrates higher hindcast skill in predicting the occurrence of HWs compared to the FCN2, especially for hindcasts initialized on lead day 0 and lead day 1. Whereas the FCN2 outperforms the AIFS model for initialization on lead day 2 onwards. As the lead time increases, the FCN2 increasingly underestimates the HWs of three categories. The under-hindcasts are attributed to persistent cold bias and weak descent over the SEA. To address this issue, a regression model is developed based on the relationship between the tropospheric temperature bias and the HW hindcast bias. The calibrated hindcasts exhibit substantial improvements, with higher threat score (TS) and equivalent threat score (ETS) than the uncalibrated hindcasts. The bias score (BS) of calibrated hindcasts maintain around 1 for all HW categories, suggesting a more balanced ratio between over-hindcasts and under-hindcasts. The regression approach outperforms the quantile mapping across these skill metrics. The findings demonstrate the efficacy of physics-based bias correction method in hindcasting the SEA HWs.</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

Improving the hindcast of late spring-early summer heatwaves in Southeast Asia with machine learning models

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Shu Gui,
  • Zihan Yang,
  • Haojie Wu,
  • Ruowen Yang,
  • Zizhen Dong,
  • Yali Yang,
  • Yu Lian

摘要

The risk of heatwaves (HWs) in the late spring–early summer period keeps escalating in Southeast Asia (SEA). To mitigate heatwave (HW) risks effectively, more accurate and longer-term forecasts are urgently needed. This study evaluates two machine learning weather prediction (MLWP) models, the FourCastNet v2 (FCN2) and the Artificial Intelligence Forecasting System (AIFS), in hindcasting the occurrence of HWs in SEA. The SEA HWs are classified into three categories using the K-means clustering: Category 1 for uniform warming across the SEA, Category 2 for pronounced warming over the Maritime Continent (MC), and Category 3 for intensified warming over the mainland SEA. Each category is driven by anomalous descent that promotes air temperature rises in their respective area. The AIFS demonstrates higher hindcast skill in predicting the occurrence of HWs compared to the FCN2, especially for hindcasts initialized on lead day 0 and lead day 1. Whereas the FCN2 outperforms the AIFS model for initialization on lead day 2 onwards. As the lead time increases, the FCN2 increasingly underestimates the HWs of three categories. The under-hindcasts are attributed to persistent cold bias and weak descent over the SEA. To address this issue, a regression model is developed based on the relationship between the tropospheric temperature bias and the HW hindcast bias. The calibrated hindcasts exhibit substantial improvements, with higher threat score (TS) and equivalent threat score (ETS) than the uncalibrated hindcasts. The bias score (BS) of calibrated hindcasts maintain around 1 for all HW categories, suggesting a more balanced ratio between over-hindcasts and under-hindcasts. The regression approach outperforms the quantile mapping across these skill metrics. The findings demonstrate the efficacy of physics-based bias correction method in hindcasting the SEA HWs.