Open AccessMachine learning in calibrating tropical cyclone intensity forecast of ECMWF EPS
Author(s) -
Chan Ming Hei Kenneth,
Wong Wai Kin,
AuYeung Kin Chung
Publication year - 2021
Publication title -
meteorological applications
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.672
H-Index - 59
eISSN - 1469-8080
pISSN - 1350-4827
DOI - 10.1002/met.2041
Subject(s) - tropical cyclone , tropical cyclone forecast model , meteorology , environmental science , numerical weather prediction , quantitative precipitation forecast , intensity (physics) , global forecast system , forecast verification , probabilistic logic , computer science , range (aeronautics) , ensemble forecasting , percentile , climatology , forecast skill , artificial intelligence , mathematics , precipitation , statistics , geography , geology , physics , quantum mechanics , materials science , composite material
Abstract Intensity prediction of tropical cyclones (TC) has been one of the major challenges for the operational forecast and warning service, as well as consequential assessment of impacts including high winds, storm surge and heavy rainfall caused by TC. With the advances in global numerical weather prediction (NWP) modelling systems, TC track and intensity forecasts for medium range are available every 6 or 12 h, and ensemble prediction system (EPS) outputs provide various scenarios for producing probabilistic forecasts. The TC intensity forecast from the EPS of the European Centre for Medium‐Range Weather Forecasts (ECMWF) has shown systematic negative biases, although the performance is better than other global models in general. A machine learning model based on XGBoost, a decision‐tree‐based machine learning algorithm, is introduced in this paper to post‐process ECMWF EPS outputs and generate an improved forecast of TC intensity. The predictors such as selected percentiles of ensemble members in maximum wind and minimum pressure of previous TC cases were applied in the XGBoost model to generate a calibrated forecast for the maximum wind of TCs. Verification of the XGBoost model was made using TCs over the western North Pacific during 2016–2019. It is found that the negative biases of the intensity forecast from ECMWF EPS and HRES can be reduced with improvement in the overall accuracy.