Weather Prediction with Diffusion Guided by Realistic Forecast Processes

TL;DR

The paper presents a conditional diffusion framework for weather forecasting that unifies direct and iterative outputs and can incorporate guidance from NWP, persistence, and climatology without retraining. By treating the lead time as a class label and using classifier-free guidance, the model learns from both unconditional and conditional distributions to produce realistic forecasts across multiple lead times. Empirical results on WeatherBench ERA5 demonstrate competitive RMSE/ACC relative to baselines, with notable gains in long-range stability when guided by realistic forecast processes. The approach offers a flexible, low-resource path toward trustworthy DL-based weather forecasting with the potential to integrate additional forecast guidance in practical workflows.

Abstract

Weather forecasting remains a crucial yet challenging domain, where recently developed models based on deep learning (DL) have approached the performance of traditional numerical weather prediction (NWP) models. However, these DL models, often complex and resource-intensive, face limitations in flexibility post-training and in incorporating NWP predictions, leading to reliability concerns due to potential unphysical predictions. In response, we introduce a novel method that applies diffusion models (DM) for weather forecasting. In particular, our method can achieve both direct and iterative forecasting with the same modeling framework. Our model is not only capable of generating forecasts independently but also uniquely allows for the integration of NWP predictions, even with varying lead times, during its sampling process. The flexibility and controllability of our model empowers a more trustworthy DL system for the general weather community. Additionally, incorporating persistence and climatology data further enhances our model's long-term forecasting stability. Our empirical findings demonstrate the feasibility and generalizability of this approach, suggesting a promising direction for future, more sophisticated diffusion models without the need for retraining.

Figures (29)

• Figure 1: Illustration of the proposed three types of guidance that are relevant to weather forecasting using diffusion models. Our model is capable of incorporating all three types of guidance to facilitate more realistic weather forecasting with diffusion models.
• Figure 2: The overall diffusion model network architecture.
• Figure 3: Generate weather forecasts from guidance of various types with our framework.
• Figure 4: RMSE (a) and ACC (c) of 500 hPa geopotential and RMSE (b) and ACC (d) of 850 hPa temperature for different baselines at 5.625$^{\circ}$ resolution. Solid lines are iterative forecasts, while dots represent direct forecasts.
• Figure 5: RMSE (a) and ACC (c) of 500 hPa geopotential and RMSE (b) and ACC (d) of 850 hPa temperature for incorporating IFS T42 and IFS T63 forecasts as guidance with different $t_0$ The dashed line is the basic 6-hourly DM iterative model, the magenta dot represents IFS T42 and the blue dot represents IFS T63.