| 摘要: |
| 数值模式预报在气象预测中占据核心地位，利用深度学习技术提升预报的空间精度是空间降尺度的重要发展方向之一。然而，传统生成式模型存在计算效率低、空间纹理精度不足以及模式坍塌等问题。为了解决这些问题，提出了一种基于残差移位扩散模型的数值模式预报空间降尺度方法。具体而言，将残差引入扩散模型的扩散过程中，帮助模型快速收敛，减少采样步骤，相比潜空间扩散模型，推理时间缩短到四十分之一，有效提高了计算效率。通过扩散模型逐步去噪的过程，生成高质量和细节丰富的降尺度图像，显著提升了空间细节表现力。此外，设计了一个噪声灵活控制机制，有效缓解了模式坍塌问题，保证了生成图像的多样性。在与生成对抗模型的对比试验中，本方法在峰值信噪比(PSNR)、结构相似性(SSIM)和图像感知相似度(LPIPS)等指标上分别实现了 6%、12% 和 16% 的显著提升。通过这一方法，有望为气象敏感行业提供更准确、更精细的天气信息，推动相关领域的技术进步。 |
| 关键词: 数值模式预报 空间降尺度 扩散模型 噪声控制 残差移位 |
| DOI：10.16032/j.issn.1004-4965.2024.076 |
| 分类号: |
| 基金项目: |
|
| A Spatial Downscaling Method for Numerical Model Prediction Based on Residual Shift Diffusion Model |
|
WANG Xing1, YE Weiliang1, ZHANG Tong2, MIAO Zishu3, WU Qiliang4
|
|
1. School of Computer Science, Nanjing University of Information Science & Technology, Nanjing 210044, China;2. School of Atmospheric Sciences, Nanjing University of Information Science & Technology, Nanjing 210044, China;3. Baotou Meteorological Bureau, Baotou, Inner Mongolia Autonomous Region 014030, China;4. Beijing Fengyun Meteorological Science and Technology Development Co., Ltd., Beijing 100081, China
|
| Abstract: |
| Numerical weather prediction (NWP) plays a central role in meteorological forecasting, and enhancing spatial resolution using deep learning techniques is a key direction in spatial downscaling. However, traditional generative models often suffer from low computational efficiency, insufficient spatial texture accuracy, and mode collapse. To address these issues, this paper proposed a spatial downscaling method for NWP based on a residual shift diffusion model. Specifically, we introduced residuals into the model’s diffusion process, aiding in quicker convergence and reducing the number of sampling steps. Unlike latent space diffusion models, our approach decreases inference time by a factor of 40, thereby significantly enhancing computational efficiency. The step-by-step denoising process of the diffusion model generated high-quality, detail-rich downscaled images, thereby enhancing the representation of spatial details. Furthermore, we designed a flexible noise control mechanism that effectively mitigated mode collapse, thereby ensuring the diversity of the generated images. In comparative experiments with generative adversarial models, our method significantly improved peak signal-to-noise ratio, structural similarity, and learned perceptual image patch similarity by 6%, 12%, and 16%, respectively. This approach promises to provide more accurate and detailed weather information for weather-sensitive industries, driving technological advancements in related fields. |
| Key words: spatial downscaling diffusion model noise control residual shift numerical weather prediction |
