Scholarlink

<p>Accurate short-term forecasting of total cloud cover (TCC) is essential for renewable energy management and atmospheric modeling but remains challenging due to complex spatiotemporal cloud dynamics. This study evaluates four deep learning (DL) architectures (Graph Attention Network, Spatio-Temporal Transformer, Graph Convolutional Network with CNN, and one-dimensional CNN) against an ARIMAX baseline for hourly TCC forecasts up to 12 hours ahead. We use CERES SYN satellite observations and ERA5 reanalysis data with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1^{\circ } \times 1^{\circ }\)</EquationSource> </InlineEquation> spatial resolution and hourly temporal frequency for 2015 to 2019 over the Texas Gulf Coast (36 grid points) and 2014 to 2019 over CONUS (2100 grid points). The Graph Attention Network (GAT) model achieves clear quantitative gains over ARIMAX, reducing RMSE by 15 to 18% and MAE by 14.5% (Regional) and 15.6% (CONUS), while improving the Index of Agreement by 11.3% and 26.5%. Correlation increases by up to 0.12 (R = 0.87 vs. 0.75) at longer leads. The novelty of this approach lies in the use of graph-based learning to capture spatial dependencies among grid cells, surpassing the localized feature extraction of CNNs. The GAT preserves diurnal cloud cycles (R &gt;0.9) and accurately tracks rapid cloud transitions (&gt;0.7 per hour). SHAP analysis identifies cloud persistence, upper-level geopotential height (400 hPa), and mid-tropospheric relative humidity (700 hPa) as dominant predictors. These results demonstrate that graph-based DL provides reproducible and quantitatively superior short-term cloud forecasts across regional and continental scales.</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

Short-Term Forecasting of Total Cloud Cover over the United States and the Texas Gulf Coast with Graph-Based Deep Learning

  • 期刊论文

  • 发表日期：

    2025年12月16日

查看全文
  • Hadi Zanganeh Kia,
  • Yunsoo Choi,
  • Rashik Islam,
  • Rijul Dimri,
  • Shihab Ahmad Shahriar

摘要

Accurate short-term forecasting of total cloud cover (TCC) is essential for renewable energy management and atmospheric modeling but remains challenging due to complex spatiotemporal cloud dynamics. This study evaluates four deep learning (DL) architectures (Graph Attention Network, Spatio-Temporal Transformer, Graph Convolutional Network with CNN, and one-dimensional CNN) against an ARIMAX baseline for hourly TCC forecasts up to 12 hours ahead. We use CERES SYN satellite observations and ERA5 reanalysis data with \(1^{\circ } \times 1^{\circ }\) spatial resolution and hourly temporal frequency for 2015 to 2019 over the Texas Gulf Coast (36 grid points) and 2014 to 2019 over CONUS (2100 grid points). The Graph Attention Network (GAT) model achieves clear quantitative gains over ARIMAX, reducing RMSE by 15 to 18% and MAE by 14.5% (Regional) and 15.6% (CONUS), while improving the Index of Agreement by 11.3% and 26.5%. Correlation increases by up to 0.12 (R = 0.87 vs. 0.75) at longer leads. The novelty of this approach lies in the use of graph-based learning to capture spatial dependencies among grid cells, surpassing the localized feature extraction of CNNs. The GAT preserves diurnal cloud cycles (R >0.9) and accurately tracks rapid cloud transitions (>0.7 per hour). SHAP analysis identifies cloud persistence, upper-level geopotential height (400 hPa), and mid-tropospheric relative humidity (700 hPa) as dominant predictors. These results demonstrate that graph-based DL provides reproducible and quantitatively superior short-term cloud forecasts across regional and continental scales.