Projected changes of typhoon intensity in a regional climate model: Development of a machine learning bias correction scheme

Abstract

A machine learning‐based bias correction scheme was developed to adjust the simulated Western North Pacific typhoon intensity in a 25‐km regional climate model (RCM). The bias correction scheme, MLERA, consists of a hybrid neural network, which takes modelled atmospheric and oceanic conditions near the storm centre as input. We use air temperature, specific humidity, and relative vorticity at 300 hPa, geopotential height at 700 hPa, wind speed at 850 hPa, sea‐level pressure, 10‐m wind speed, and total air‐sea fluxes. Those predictors are selected using least absolute shrinkage and selection operator (Lasso) algorithm and principal component analysis (PCA). Because there is no ‘ground truth’ for RCM simulated storms, we train and test MLERA using ERA‐Interim reanalysis and best‐track data. The intensity statistics estimated from MLERA match well with those from observations. MLERA, when applied to RCM history base period, increases the likelihood of simulated typhoons (with wind speed >33 m/s) from 40.8 to 73.9% in the direct simulation. Meanwhile, the RCM itself produces no C3+ storms (wind speed >50 m/s) in both RCP4.5 and RCP8.5 scenarios, while MLERA significantly increases those storms. MLERA suggests an increasing trend of the frequency of violent typhoons from near‐term (2031–2060) to long‐term (2069–2098). Such an increase is consistent with our understanding that a warming climate will increase the storm intensity. The present study shows the potential of the machine learning technique for bias correcting cyclone intensity in RCMs. Furthermore, the proposed algorithm could also be applied to the next generation of high‐resolution global climate models, which may have a spatial grid spacing close to today's RCMs.

Application of XGBoost and TrajGRU to improve the accuracy of ECMWF wind forecasts

“…In the past, many studies have proposed many post-processing methods to correct the bias of NWP. The MLERA scheme was designed using machine learning methods to correct the bias of typhoon forecasts in the WNP using multiple sea surface variables, which improved the forecast accuracy [4]. Duan et al used a recurrent neural network to achieve short-term wind speed prediction and used error decomposition technology to further improve the prediction accuracy [5].…”

Comparison of Bias Correction Methods for Summertime Daily Rainfall in South Korea Using Quantile Mapping and Machine Learning Model

“…Machine learning has an advantage in identifying meaningful information in the climate system through pattern recognition and feature extraction techniques, which eventually help solve the problems of nonlinear phenomenon prediction [27][28][29][30]. This suggests that machine learning can be used in the bias correction of climate model data [25,[31][32][33][34][35][36]. Kim et al [25] utilized the LSTM machine learning model as a bias correction method to improve MJO forecasts.…”