Autoregressive Atmospheric Profile Data Assimilation using Diffusion Model

Weather forecasting is a critical component in numerous scientific and practical domains, such as aviation, agriculture, and disaster management. To provide accurate weather predictions, we propose an autoregressive data assimilation framework that is capable of real-world applications. This framework integrates deep learning-based diffusion models to assimilate GK2A AAP and radiosonde in real time. By doing so, it bridges the gap between data assimilation and predictive modeling, offering a robust solution for accurate high-resolution weather forecasting. The reanalysis field ERA5 was used as a background for training, but in the inference process, except for the initial state, the results predicted by the deep learning model were all used as the background for the next step. The model was validated against unassimilated radiosonde data with temperature and relative humidity for 10 pressure levels: 100, 150, 200, 250, 400, 500, 700, 850, 925, and 1000 hpa in 2021 and 2022. The results show that the model has higher sonde accuracy in the lower layers than ERA5, while ERA5 has higher accuracy in the upper layers. This is because the spatial distribution of GK2A observations is different from that of ERA5 in the upper layers. However, as the data assimilation period increased, the spatial distribution was similar to that of ERA5. Our model enhanced the precision and practical applicability of weather prediction models, demonstrating robustness and operational viability under real-world conditions.