Bias correction of tropical cyclone intensity for ensemble forecasts using the XGBoost method

In this study, the extreme gradient boosting (XGBoost) algorithm is used to correct tropical cyclone (TC) intensity in ensemble forecast data from the Typhoon Ensemble Data Assimilation and Prediction System (TEDAPS) at the Shanghai Typhoon Institute (STI), China Meteorological Administration (CMA). Results show that the forecast accuracy of TC intensity may be improved substantially using the XGBoost algorithm, especially when compared with a simple ensemble average of all members in the ensemble forecast [as depicted by the ensemble average (EnsAve) algorithm in this study]. The forecast errors for maximum wind speed (MWS) and minimum sea-level pressure (MSLP) have been reduced by a significant margin, ranging from 6.3% to 18.4% for MWS and from 4% to 14.9% for MSLP, respectively. The performance of the XGBoost algorithm is overall better than that of the EnsAve algorithm, although there are a few samples when it is worse. The bias analysis shows that TEDAPS underpredicts the MWS and overpredicts the MSLP, meaning that the TEDAPS underestimates TC intensity. However, the XGBoost algorithm can reduce the bias to improve the forecast accuracy of TC intensity. Specifically, it achieves a reduction of over 20% in forecast errors for both the MWS and MSLP of typhoons compared to the EnsAve algorithm, indicating the XGBoost algorithm’s particular advantage in forecasting intense TCs. These results indicate that the TC intensity forecast can be substantially improved using the XGBoost algorithm, relative to the EnsAve algorithm.