Scholarlink

<p>The boundary-layer jet (BLJ) over the northern South China Sea exerts a strong influence on heavy rainfall along the South China coast, yet its predictability across different modeling frameworks remains insufficiently understood. This study presents a systematic comparison of BLJ forecasts from a conventional numerical weather prediction (NWP) model (the European Centre for Medium-Range Weather Forecasts, ECMWF) and a data-driven artificial intelligence (AI) model (Pangu). The AI model outperforms the NWP model in overall accuracy, with reduced root-mean-square errors and improved representation of wind direction. However, the two models exhibit distinct and opposite systematic error structures. ECMWF consistently overestimates BLJ intensity and spatial extent, accompanied by a northwestward displacement of the jet core and a clockwise wind-direction bias. Conversely, Pangu underestimates BLJ intensity, with a northeastward core shift and a counterclockwise wind-direction bias. Motivated by these complementary errors, we develop a U-Net–based blending framework with a composite loss function integrating mean absolute error, miss rate, and false alarm rate. The blended forecasts significantly outperform both individual models across multiple BLJ metrics. These results demonstrate that combining physics-based and AI-based forecasts can mitigate systematic biases to improve BLJ prediction, highlighting a promising pathway toward hybrid forecasting systems for weather extremes.</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

Complementary error structures of AI and numerical models in forecasting boundary-layer jets over the South China Sea

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月22日

查看全文
  • Haimin Li,
  • Yu Du

摘要

The boundary-layer jet (BLJ) over the northern South China Sea exerts a strong influence on heavy rainfall along the South China coast, yet its predictability across different modeling frameworks remains insufficiently understood. This study presents a systematic comparison of BLJ forecasts from a conventional numerical weather prediction (NWP) model (the European Centre for Medium-Range Weather Forecasts, ECMWF) and a data-driven artificial intelligence (AI) model (Pangu). The AI model outperforms the NWP model in overall accuracy, with reduced root-mean-square errors and improved representation of wind direction. However, the two models exhibit distinct and opposite systematic error structures. ECMWF consistently overestimates BLJ intensity and spatial extent, accompanied by a northwestward displacement of the jet core and a clockwise wind-direction bias. Conversely, Pangu underestimates BLJ intensity, with a northeastward core shift and a counterclockwise wind-direction bias. Motivated by these complementary errors, we develop a U-Net–based blending framework with a composite loss function integrating mean absolute error, miss rate, and false alarm rate. The blended forecasts significantly outperform both individual models across multiple BLJ metrics. These results demonstrate that combining physics-based and AI-based forecasts can mitigate systematic biases to improve BLJ prediction, highlighting a promising pathway toward hybrid forecasting systems for weather extremes.