Optimum Scenario Selection Using Ensemble Clustering and Bayesian Estimation of Realization Probability: Application to Typhoon Hagibis (2019)

We describe a novel method for objectively selecting the most probable scenario among four typhoon forecasts. The method uses ensemble clustering to progressively incorporate a sequence of analytical data leading up to the most recent. Each ensemble member and the analytical data were initially projected onto the phase space spanned by the two leading principal components from an empirical orthogonal function analysis of the ensemble clustering. We then employed a particle filter-based Bayesian approach to assess the similarity between the forecasts and the analytical results within phase space. The scenario with the highest probability was then selected as the optimal scenario by application of the selective ensemble method. This new method was applied to Typhoon Hagibis (2019) using the regional ensemble prediction system of the Japan Meteorological Agency (JMA). The selected scenario successfully predicted a mesoscale front and associated coastal heavy rainfall that exceeded 100 mm per 3 hours in Miyagi Prefecture, which JMA's deterministic mesoscale model failed to forecast. Notably, the optimal scenario was identified prior to the onset of heavy rainfall. Statistical analysis of multiple typhoon cases demonstrated that the proposed method allows for optimal scenario selection up to six hours in advance of the target time. These results suggest that the new method can identify more realistic scenarios than operational deterministic forecasts of significant weather phenomena before their occurrence.