Application of Best-Subsets Multiple Linear Regression Models in Forecasting the Gale-Force Wind Radii of Tropical Cyclones

Based on the International Best Track Archive for Climate Stewardship dataset and European Centre for Medium-Range Weather Forecasts reanalysis data, the best-subsets multiple linear regression (bs-MLR) models were established by forecasting the gale-force wind radii (R17) of Tropical Cyclones (TCs) in the western North Pacific region. First, TCs from June to November 2001–2014 were divided into four categories according to the 1−25, 26−50, 51−75, and 76−100 percentiles of the initial sizes (R_{17_0}), and the bs-MLR models for TCs in each category were established. Then all TCs from June to November 2015 were used to test the estimated effectiveness of the bs-MLR models. The results showed that, when R_{17_0} was less than 92.6 km or R_{17_0} was between 111.1 km and 138.9 km, the models had better performances in forecasting the values and changing tendencies of R17 in the next 12 h (R_{17_12}) for any moment of TC life cycle. When R_{17_0} was between 111.1 km and 138.9 km, the models had better performances in forecasting the values and changing tendencies of R17 in the next 24 h (R_{17_24}) for any moment of TC life cycle. Overall, bs-MLR models had higher accuracy in forecasting R_{17_12} than R_{17_24}.