基於擴散模型的短期天氣預報

Abstract

高空間解析度的區域天氣預報對於災害防治與資源管理至關重要，但對數據驅動（Data-driven）模型而言，這仍是一項重大挑戰。由於高解析度模型對計算資源的需求極高，受限於硬體運算能力的上限，實務上往往必須將預報範圍限縮至特定的局部空間窗口。然而，儘管全球機器學習預報模型已取得卓越成效，但若將其直接應用於有限區域（Limited-area）時，模型感受野（Receptive fields）與局部物理尺度往往無法匹配，導致細微特徵流失及空間不穩定性。本研究提出了一套專為區域領域設計的擴散預報模型（Diffusion-based forecasting models）的適應性框架。研究證實，針對全球尺度設計的傳統深層模型架構在應用於0.25◦ 解析度的網格時，容易產生過度平滑（Over-smoothing）現象。為解決此問題，我們提出一種局部適應策略，結合較淺層的U-Net 架構、靜態地理條件約束（Static geographic conditioning），以及包含頻譜（Spectral）與梯度損失函數在內的一系列輔助訓練目標。在台灣複雜地形區域的實驗結果顯示，該適應框架在機率預報能力（Probabilistic skill）上有顯著提升。具體而言，與未經適應的基準模型相比，本模型在2 公尺氣溫及10 公尺風場分量等近地表變量中，展現出更低的連續分級機率評分（CRPS）。定性分析進一步證實，該框架能有效重建精確的地形氣溫梯度，並保留高頻空間紋理。儘管區域模型與全球尺度系統之間仍存在差距，本研究為提升數據驅動區域系集預報（Ensemble forecasting）的物理一致性與預報準確度，提供了一個穩健且具可行性的研究方向。

Localized weather forecasting at high spatial resolutions is critical for disaster mitigation and resource management, yet it remains a significant challenge for data-driven models. Due to the substantial computational overhead of high-resolution models, hardware limitations often necessitate the restriction of the forecasting domain to a fixed spatial window. However, while global machine learning forecasting models have achieved remarkable success, their direct application to limited-area regional domains often suffers from a mismatch between model receptive fields and local physical scales, leading to the loss of fine-scale details and spatial instabilities. This study proposes a specialized adaptation framework for diffusion-based forecasting models tailored for regional domains. We demonstrate that traditional deep architectures designed for global scales can lead to oversmoothing when applied to smaller grids at 0.25◦ resolution. To address this, we propose a localized adaptation strategy that combines a shallower U-Net architecture with static geographic conditioning and a suite of auxiliary training objectives, including spectral and gradient-based losses. Our experimental results, conducted over the complex topography of the Taiwan region, indicate that the proposed adaptation yields measurable improvements in probabilistic skill. Specifically, the model achieves lower Continuous Ranked Probability Scores (CRPS) for near-surface variables such as 2-meter temperature and 10- meter wind components compared to unadapted baselines. Qualitative analysis further confirms that the framework effectively reconstructs sharp orographic temperature gradients and preserves high-frequency spatial textures. While a gap remains between regional models and global-scale systems, this work provides a robust pathway for enhancing the physical consistency and predictive accuracy of data-driven regional ensemble forecasting.