Ensemble seasonal forecasting of typhoon frequency over the western North Pacific using multiple machine learning algorithms

Abstract

This study introduces an ensemble prediction methodology employing multiple machine learning algorithms for forecasting the frequency of typhoons (TYFs) over the western North Pacific (WNP) during June‒November. Potential predictors were initially identified based on the relationships between the year-by-year variation (DY) of the TYFs and preseason (March–May) environmental factors. These predictors were subsequently further refined, resulting in the selection of eight key predictors. Prediction models were constructed using twenty machine learning algorithms, utilizing data from 1965 to 2010. These trained models were then applied to perform hindcasts of TYFs from 2011 to 2023. The forecasted DY was added to the observed TYF of the preceding year to obtain the current year’s TYF. The results indicate that the TYFs predicted by the multi-model ensemble (MME) closely align with the observation during the hindcast period. Compared to individual models, the MME improves the prediction skill for the DY by at least 5.56% and up to 56.92%. Furthermore, the mean bias of the MME for TYF is notably smaller than that of the ECMWF’s most recent seasonal forecasting system (SEAS5) in the years of 2017‒2023. The superior performance of the ensemble prediction approach was also validated through leave-one-out cross-validation. This research underscores the potential of ensemble prediction approach utilizing multiple machine learning algorithms to improve the forecasting skill of TYF over the WNP.

Keywords

• typhoon frequency
• seasonal prediction
• machine learning
• multi-model ensemble