Scholarlink

Abstract <p>NWP models have difficulty in converting complex microphysical and dynamical processes in lightning. In contrast, data-driven AI models fail to capture the latent features associated with lightning activity. A novel NWP–AI hybrid lightning early warning system using a two-autoencoder-based classification model (<i>C</i><sub><i>2AE</i></sub>), which uses IMD WRF 9 km forecast as input, is discussed. The current version of <i>C</i><sub><i>2AE</i></sub> uses mean square error as a loss function and, after finetuning, can forecast lightning activity with an error of 3% when tested over the training region. Further analysis of different thunderstorm-prone regions for March–April–May 2020 reveals that <i>C</i><sub><i>2AE</i></sub> can capture the spatial and temporal distribution of lightning independent of the training region.</p> Research highlights <p><UnorderedList Mark="Bullet"> <ItemContent> <p>A two-autoencoder-based classification model (<i>C</i><sub><i>2AE</i></sub>) architecture is described.</p> </ItemContent> <ItemContent> <p>The model, when tested on the training domain, has statistically significant skills with a Bayes error of only 3%.</p> </ItemContent> <ItemContent> <p>Further analysis of model performance over different thunderstorm-prone regions for MAM 2020 shows the model is able to capture the spatial and temporal distribution of lightning activity relatively well.</p> </ItemContent> </UnorderedList></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

Performance of a novel NWP–AI hybrid lightning early warning system over Indian Subcontinent

  • 期刊论文

  • 发表日期：

    2025年1月27日

查看全文
  • Rituparna Sarkar,
  • Parthasarathi Mukhopadhyay,
  • Sunil D Pawar

摘要

Abstract

NWP models have difficulty in converting complex microphysical and dynamical processes in lightning. In contrast, data-driven AI models fail to capture the latent features associated with lightning activity. A novel NWP–AI hybrid lightning early warning system using a two-autoencoder-based classification model (C2AE), which uses IMD WRF 9 km forecast as input, is discussed. The current version of C2AE uses mean square error as a loss function and, after finetuning, can forecast lightning activity with an error of 3% when tested over the training region. Further analysis of different thunderstorm-prone regions for March–April–May 2020 reveals that C2AE can capture the spatial and temporal distribution of lightning independent of the training region.

Research highlights

A two-autoencoder-based classification model (C2AE) architecture is described.

The model, when tested on the training domain, has statistically significant skills with a Bayes error of only 3%.

Further analysis of model performance over different thunderstorm-prone regions for MAM 2020 shows the model is able to capture the spatial and temporal distribution of lightning activity relatively well.