Improving the hindcast of the northward shift of South Asian high in June with machine learning

Abstract

Accurate prediction of the northward shift of the South Asian High (SAH) in June is crucial for improving the flood and drought management of Asian countries during the summer. This study investigates the ability of three supervised machine learning (ML) models in predicting the meridional index of the SAH (SAHI) in June. The ML models include the extreme gradient boosting (XGBoost), the support vector machine (SVM), and the multi-layer perceptron (MLP) neural network. The training data is derived from the Coupled Model Intercomparison Project Phase 6 (CMIP6) model data that is significantly correlated with the reference data at a 99% confidence level. The hyperparameter optimization (HPO) is performed for each ML model using the particle swarm optimization (PSO). Six objective functions are defined for the HPO based on the conventional root-mean-square error (RMSE), the interannual variability skill score, and the temporal correlation coefficient (TCC). The performance of optimized ML models is evaluated with the TCC and the same sign rate (SSR). The top two models are the PSO-XGBoost model tuned with RMSE+IVS and the PSO-SVM model tuned with log(RMSE+TCC). Their stacked ensemble model has the TCC of 0.54 and the SSR of 72%. The average of the best model hindcasts has a higher TCC of 0.61 but a lower SSR of 67% than the ensemble model. Further investigation suggests that the ensemble model only preserves the predictor-predictand relationships for two predictors. To improve the representation of the predictor-predictand relationship, we divided the predictors into two groups and trained the ML models separately with interannual increments of predictors in Group 1 and standardized anomalies of predictors in Group 2. The average of the best model hindcasts from the two groups have the TCC of 0.63 and the SSR of 72%. The improvement in the SAHI hindcast is associated with a more realistic predictor-predictand relationship in the ML models. The hindcast results are comparable to the operational models and thus confirm the extensive application potential of the ML models in the SAH prediction.