[M-GI27] Data-driven approaches for weather and hydrological predictions

[MGI27-P03] Probabilistic Ensemble Generation Using a Diffusion Model Trained on JMA MSM Data

*Natsumi Saito¹, Daiya Shiojiri², Shunji Kotsuki² (1.Chiba University, Graduate School of Science and Engineering, 2.Chiba University, Institute for Advanced Academic Research)

Keywords:diffusion model, ensemble prediction, MSM, MEPS

In recent years, heavy rainfall disasters have become increasingly severe in Japan. Here ensemble prediction systems (EPS) are useful sources of probabilistic information for predicting potential hazards, such as the location and intensity of rainfall. To enhance the detection of potential precipitation-induced hazards, the Japan Meteorological Agency (JMA) operates a regional EPS called the Mesoscale Ensemble Prediction System (MEPS), which aims at quantitatively evaluating uncertainties in deterministic forecasts produced by the Mesoscale Model (MSM). However, due to limited computational resources, MEPS is updated less frequently than MSM, and the ensemble size is also restricted to be 21.
To address the issues of update frequency and ensemble size of ensemble weather predictions, this study explores the feasibility of using a generative deep learning model that is both fast and capable of generating ensemble predictions. Specifically, we use a diffusion model, a kind of generative deep learning model known for its superior learning stability compared to conventional deep-learning models. We train the diffusion model on MSM forecast state variables to generate probabilistic ensemble predictions that resemble MSM outputs. The characteristics of the generated ensembles are evaluated through comparison with MEPS. This presentation will provide the most recent results at the time of the conference.