Scholarlink

<p>High-fidelity synthetic meteorological data are critical for building energy simulation and climate resilience studies. However, existing generative models often exhibit fitting bias, limited temporal coherence, and reduced generalization across heterogeneous climatic regions. To address these limitations, this study proposes a High-Fidelity Synthetic Meteorological Year (HF-SMY) generation framework that integrates channel attention, physics-guided spatial attention, and multiple physics-informed constraint losses. The model incorporates meteorological and statistical priors into a unified generative architecture to improve physical consistency and temporal structure while enhancing the representation of extreme events. Experimental evaluation shows that the proposed framework improves the accuracy of key meteorological variables, with mean absolute errors of 1.1&#xa0;hPa for pressure, 0.78 ℃ for temperature, and 0.45&#xa0;m/s for wind speed. Compared with variational autoencoder (VAE) and generative adversarial network (GAN) baselines, it achieves a 38% reduction in normalized mean absolute error (NMAE). The model also better preserves diurnal cycles and temporal continuity patterns. Statistical tests further indicate that the generated distributions of heavy rainfall intensity and high-temperature frequency are consistent with observations, with no significant differences at the 0.05 level. In addition, results suggest improved robustness across climatic regions, limiting performance variability between coastal and inland environments to below 1%. These findings indicate that incorporating physical priors into deep generative models can improve the realism and generalization of synthetic meteorological data, supporting downstream applications in energy system modeling and climate risk assessment.</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

High-fidelity synthetic meteorological year generation model integrating physical priors and dual attention

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Rong Hu,
  • Xinyi Chen,
  • Fei Du

摘要

High-fidelity synthetic meteorological data are critical for building energy simulation and climate resilience studies. However, existing generative models often exhibit fitting bias, limited temporal coherence, and reduced generalization across heterogeneous climatic regions. To address these limitations, this study proposes a High-Fidelity Synthetic Meteorological Year (HF-SMY) generation framework that integrates channel attention, physics-guided spatial attention, and multiple physics-informed constraint losses. The model incorporates meteorological and statistical priors into a unified generative architecture to improve physical consistency and temporal structure while enhancing the representation of extreme events. Experimental evaluation shows that the proposed framework improves the accuracy of key meteorological variables, with mean absolute errors of 1.1 hPa for pressure, 0.78 ℃ for temperature, and 0.45 m/s for wind speed. Compared with variational autoencoder (VAE) and generative adversarial network (GAN) baselines, it achieves a 38% reduction in normalized mean absolute error (NMAE). The model also better preserves diurnal cycles and temporal continuity patterns. Statistical tests further indicate that the generated distributions of heavy rainfall intensity and high-temperature frequency are consistent with observations, with no significant differences at the 0.05 level. In addition, results suggest improved robustness across climatic regions, limiting performance variability between coastal and inland environments to below 1%. These findings indicate that incorporating physical priors into deep generative models can improve the realism and generalization of synthetic meteorological data, supporting downstream applications in energy system modeling and climate risk assessment.