Contribution of lagged members to the performance of a global ensemble prediction system

Abstract

National Centre for Medium Range Weather Forecasting (NCMRWF), India runs operationally a global ensemble prediction system (NEPS-G) which has 23 ensemble members. The control and 11 perturbed ensemble members have the initial conditions of 00 UTC. The start time of the other 11 perturbed members is 12 UTC of the previous day. The objective of the present paper is to study the effect of the lagged ensemble members on the forecast quality of NEPS-G. For this, we have carried out an inter-comparison of the relative skills of (i) 11-member ensemble starting from 00 UTC i.e., E00_11, (ii) 11 lagged ensemble members starting from 12 UTC i.e., E12_11 and the (iii) operational 23-member ensemble i.e. E23 using the verification of temperature, wind and geopotential height forecasts for 7-day forecast lead times for April–May-June 2019 (AMJ) of summer and December 2019–January 2020 (DJ) of winter. The temperature at 850 hPa and geopotential height at 500 hPa are verified over the northern hemisphere. The zonal and meridional winds at 850 hPa and 200 hPa are verified over the region 0oN-40oN, 60°E-110°E. The metrics used for validation are ensemble-spread and root-mean-square error (RMSE) relationship, brier skill score (BSS), reliability diagram, outliers statistics, relative operating characteristics (ROC) score and ranked probability score (RPS). For brevity, only the results for Z500, T850 and U850 have been discussed in detail here.The results show that in both seasons, the RMSE-spread relationship is better for E23 than for E00_11 and E12_11. For T850 and Z500 over NH, both E23 and E00_11 have higher RMSE and higher ensemble spread in DJ season as compared to AMJ season. In case of wind over the Indian region, both RMSE and ensemble spread exhibit lower values during DJ as compared to AMJ for both ensemble configurations. For both seasons, the BSS of E23 is positive and the ROC score is >0.5 beyond 7 days forecast times for all the variables for all three ensembles and the percentage of outliers in E23 is also less. Overall, the contribution of 11 lagged ensemble members improves the skill of NEPS-G in both seasons and skill is better during winter (DJ season). The error in the mean forecast track for tropical cyclone ‘Fani’ for E23 is found to be less than for E00_11.We have also compared the forecast skill of E23 with that of the ensemble formed by 22 members, all running from 00 UTC i.e. E00_22 for the same set of variables as well as precipitation for one month forecast data from each season, namely June 2019 from summer and January 2020 from winter. Results show that the RMSE-spread relationship is very much similar for both configurations and both seasons. For E00_22 and E23, the ensemble spread is closer to the RMSE of ensemble mean for all the variables in January compared to June. BSS was slightly higher for E00_22 but was not significant at all lead time for all the variables except Z500 in January which shows significant improvement in BSS of E00_22 from day 2 to day 5. In precipitation forecast verification, RPS for E00_22 is slightly less (though significant) than E23 at all forecast lead times in June while in January there was no significant difference. So, E23 dataset of NEPS-G which demands less computational resources at the same time offers more skilful forecast than E00_11 over a longer forecast lead time and is nearly as skilful as E00_22.