Scholarlink

<p>Although deep learning (DL) models excel in deterministic weather forecasting, their application in operational decision-making is often constrained by systematic biases and a lack of uncertainty quantification. This study addresses these limitations by introducing an integrated framework that couples a CNN–GRU–MHA model with Bayesian Joint Probability (BJP) post-processing to transform deterministic monthly average temperature forecasts into calibrated probabilistic distributions. The framework is benchmarked against the established Ensemble Marginal Model Statistics (EMOS) method. Results demonstrate significant skill enhancements, with the DL-BJP framework achieving up to 63% improvement in Ranked Probability Scores (RPS) over raw DL outputs and outperforming the EMOS baseline in the majority of months. Evaluation via seasonal ROC curves shows outstanding discriminative power (AUC values ranging from 0.78 to 0.98). Furthermore, reliability analysis reveals that the framework provides accurate probabilistic outlooks for “Above Normal” and “Below Normal” categories, supporting the detection of extreme events, while the “Normal” category exhibits comparatively higher calibration sensitivity. This research highlights that the integrated approach offers a practical pathway for generating reliable, uncertainty-aware S2S predictions, providing essential intelligence for risk-based decision-making in climate-sensitive sectors.</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

An integrated deep learning-bayesian framework for probabilistic temperature prediction

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Behshid Yasavoli,
  • Arezou Habibirad,
  • Zohreh Javanshiri,
  • Mehdi Jabbari Nooghabi

摘要

Although deep learning (DL) models excel in deterministic weather forecasting, their application in operational decision-making is often constrained by systematic biases and a lack of uncertainty quantification. This study addresses these limitations by introducing an integrated framework that couples a CNN–GRU–MHA model with Bayesian Joint Probability (BJP) post-processing to transform deterministic monthly average temperature forecasts into calibrated probabilistic distributions. The framework is benchmarked against the established Ensemble Marginal Model Statistics (EMOS) method. Results demonstrate significant skill enhancements, with the DL-BJP framework achieving up to 63% improvement in Ranked Probability Scores (RPS) over raw DL outputs and outperforming the EMOS baseline in the majority of months. Evaluation via seasonal ROC curves shows outstanding discriminative power (AUC values ranging from 0.78 to 0.98). Furthermore, reliability analysis reveals that the framework provides accurate probabilistic outlooks for “Above Normal” and “Below Normal” categories, supporting the detection of extreme events, while the “Normal” category exhibits comparatively higher calibration sensitivity. This research highlights that the integrated approach offers a practical pathway for generating reliable, uncertainty-aware S2S predictions, providing essential intelligence for risk-based decision-making in climate-sensitive sectors.