Assessment of MME methods for seasonal prediction using WMO LC‐LRFMME hindcast dataset

Abstract

AbstractDifferent multi‐model ensemble (MME) methods were investigated for their potential to improve the skill of 1‐month lead seasonal forecast products, based on six models from Global Producing Centers (GPCs) for long‐range forecasts (LRFs) designated by the World Meteorological Organization (WMO). We first compared the hindcast performance of seven MME methods (simple composite method, SCM; simple linear regression, SLR; multiple linear regression, MLR; best selection anomaly, BSA; multilayer perceptron, MLP; radial basis function, RBF; genetic algorithm, GA) for the global 2‐m temperature and precipitation for 1983–2009. The reference datasets for 2‐m temperature and precipitation are the ERA‐Interim from European Centre for Medium‐Range Weather Forecasts (ECMWF) and Global Precipitation Climatology Project (GPCP) for hindcast verification. For real‐time verification, the data from the National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) reanalysis 1 for 2‐m temperature and climate anomaly monitoring system and outgoing longwave radiation precipitation index (CAMS OPI) for precipitation are used. According to our analysis, GA was the most successful MME method in predicting both the global 2‐m temperature and precipitation for all four seasons. GA also showed good performance in predicting the 2‐m temperature and precipitation over the 13 regional climate outlook forum (RCOF) regions in all four seasons, but the range in performance among the RCOF regions varied significantly. In a real‐time forecast period (MAM 2012‐DJF 2015/16), GA outperformed in terms of time‐averaged anomaly pattern correlation coefficient (ACC) and root‐mean‐square error (RMSE) of the 2‐m temperature, although the forecast skill difference (0.02) between GA and SCM was not statistically significant. For the precipitation, both SCM and GA also reveal better performance than other MME methods. During the very strong El Niño event in 2015, individual models show better performance than other years. Nonetheless, these two MME methods outperform all the individual models.