Scholarlink

<p>Deep learning, a subset of artificial intelligence (AI), facilitates processing vast amounts of data with remarkable precision. Deep learning methods utilize deep neural networks to learn and capture the patterns between data. Because of the wide range of applications of deep learning methods in various domains, the complex and nonlinear relationships among climate data series, and the crucial role that accurate temperature prediction plays in many sectors, our objective in this study is to predict the daily average temperature using both individual and hybrid deep learning methods. In this study, we aim to utilize the key characteristics of different architectures to develop an optimized hybrid model to predict average temperatures more accurately. Our suggested hybrid model (CNN-GRU-MHA) integrates CNN for spatial feature extraction, GRU for capturing both short-term and long-term dependencies, and TFT’s multi-head attention mechanism to concentrate on the most important time points, along with Bayesian optimization to ensure optimal configurations for each method. We used a dataset spanning over 30 years of minimum, average, and maximum daily temperature records from the Mashhad meteorological station. Then we assessed the predictive accuracy of each model using three different lead times, measures, and a statistical t-test. The results demonstrate that our proposed hybrid model significantly outperforms individual and other hybrid models, achieving the lowest root mean square error (RMSE) and Theil’s U2 measures and also the highest Nash–Sutcliffe model efficiency coefficient (NSE) across all lead times. This study highlights the advantages of incorporating spatial, sequential, and attention mechanisms to enhance temperature forecasting, especially for long-term predictions.</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

A hybrid deep learning model in predicting weather temperature

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Behshid Yasavoli,
  • Arezou Habibirad,
  • Zohreh Javanshiri

摘要

Deep learning, a subset of artificial intelligence (AI), facilitates processing vast amounts of data with remarkable precision. Deep learning methods utilize deep neural networks to learn and capture the patterns between data. Because of the wide range of applications of deep learning methods in various domains, the complex and nonlinear relationships among climate data series, and the crucial role that accurate temperature prediction plays in many sectors, our objective in this study is to predict the daily average temperature using both individual and hybrid deep learning methods. In this study, we aim to utilize the key characteristics of different architectures to develop an optimized hybrid model to predict average temperatures more accurately. Our suggested hybrid model (CNN-GRU-MHA) integrates CNN for spatial feature extraction, GRU for capturing both short-term and long-term dependencies, and TFT’s multi-head attention mechanism to concentrate on the most important time points, along with Bayesian optimization to ensure optimal configurations for each method. We used a dataset spanning over 30 years of minimum, average, and maximum daily temperature records from the Mashhad meteorological station. Then we assessed the predictive accuracy of each model using three different lead times, measures, and a statistical t-test. The results demonstrate that our proposed hybrid model significantly outperforms individual and other hybrid models, achieving the lowest root mean square error (RMSE) and Theil’s U2 measures and also the highest Nash–Sutcliffe model efficiency coefficient (NSE) across all lead times. This study highlights the advantages of incorporating spatial, sequential, and attention mechanisms to enhance temperature forecasting, especially for long-term predictions.