Scholarlink

<p>Extreme temperature events such as heatwaves and coldwaves represent major climate hazards that pose increasing risks to human health, water and energy systems, agriculture, and infrastructure. Accurate prediction of these rare events is therefore critical for improving environmental risk assessment and early warning systems. In this study, we develop a deep learning–based framework for predicting heatwaves and coldwaves using daily meteorological observations from five ground weather stations. Nine deep learning architectures, including recurrent, graph-based, spatio-temporal, and attention-based neural networks, are evaluated across multiple forecasting horizons. Predicting extreme temperature events presents substantial challenges due to the severe class imbalance between rare extreme events and normal climate conditions, as well as their multi-day duration. To address these challenges, we implement an optimized cost-sensitive learning strategy that prioritizes minority extreme-event classes during model training, together with a spell-based evaluation framework designed for extended climate events. Experimental results show that the spatio-temporal graph convolutional network with gated recurrent units (STGCN-GRU) and the temporal convolutional network (TCN) achieve the best predictive performance in terms of F1 score across forecasting horizons. Ablation experiments further demonstrate that the proposed cost-sensitive learning strategy provides more effective handling of class imbalance than conventional resampling approaches. The proposed framework offers a promising data-driven approach for improving the prediction and risk assessment of rare temperature 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

Deep learning prediction of rare heatwave and coldwave events under severe class imbalance

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Abdullah Fadhil Tawfeeq,
  • Ümit Haluk Atasever

摘要

Extreme temperature events such as heatwaves and coldwaves represent major climate hazards that pose increasing risks to human health, water and energy systems, agriculture, and infrastructure. Accurate prediction of these rare events is therefore critical for improving environmental risk assessment and early warning systems. In this study, we develop a deep learning–based framework for predicting heatwaves and coldwaves using daily meteorological observations from five ground weather stations. Nine deep learning architectures, including recurrent, graph-based, spatio-temporal, and attention-based neural networks, are evaluated across multiple forecasting horizons. Predicting extreme temperature events presents substantial challenges due to the severe class imbalance between rare extreme events and normal climate conditions, as well as their multi-day duration. To address these challenges, we implement an optimized cost-sensitive learning strategy that prioritizes minority extreme-event classes during model training, together with a spell-based evaluation framework designed for extended climate events. Experimental results show that the spatio-temporal graph convolutional network with gated recurrent units (STGCN-GRU) and the temporal convolutional network (TCN) achieve the best predictive performance in terms of F1 score across forecasting horizons. Ablation experiments further demonstrate that the proposed cost-sensitive learning strategy provides more effective handling of class imbalance than conventional resampling approaches. The proposed framework offers a promising data-driven approach for improving the prediction and risk assessment of rare temperature extremes.