A Machine Learning (ML)-Based Approach to Improve Tropical Cyclone Intensity Prediction of NCMRWF Ensemble Prediction System

Global numerical weather prediction (NWP) models, including ensemble prediction systems (EPS), are routinely used by the forecasters to predict tropical cyclone (TC) tracks and intensity. However, due to computational restrictions, the EPSs are usually of coarse resolution, which results in poor prediction of TC intensity. The bias in the model predicting maximum sustained winds (MSW) and central pressure (CP) is large when TCs are intense. This article describes the suitability of machine learning (ML) techniques to reduce errors in TC intensity forecasts obtained from the National Centre for Medium Range Weather Forecasting (NCMRWF) Ensemble Prediction System (NEPS) over the North Indian Ocean (NIO). Four different ML techniques, namely, Multivariate Linear Regression (MLR), Support Vector Regression (SVR), Random Forest (RF), and eXtreme Gradient Boost (XGB), have been used for the bias correction (BC) of mean MSW and CP, while the spread of ensemble members has been retained. The study is based on 20 TC cases formed over the NIO during 2018-21. The best track (BT) information from the India Meteorological Department (IMD) has been used for training the ML models and verifying them. The results show that XGB (for CP) and RF (for MSW) ML techniques are superior to MLR and SVR methods. The statistically significant reduction (at 99% CI) in mean absolute error (MAE) (root mean square error (RMSE) of mean MSW and CP while using the best algorithm are 30 (33)% and 63 (55)%, respectively. The correlation coefficient for CP increases from 0.58 to 0.93 and that for MSW from 0.60 to 0.75. The probabilistic verification of bias-corrected ensemble members also suggests that the RF- and XGB-based models have better reliability and ROC compared to the raw and other models for MSW and CP, respectively.