jma_logo Welcome to TCC wmo_logo
World Climate
Climate System Monitoring
El Niño Monitoring
NWP Model Prediction
Global Warming
Climate in Japan
Training Module
Press release
HOME > Ensemble Model Prediction > Description of JMA's Seasonal Ensemble Prediction System (JMA/MRI-CPS3)

Description of the Seasonal Ensemble Prediction System (JMA/MRI-CPS3)

JMA operates the seasonal ensemble prediction system for sub-seasonal to seasonal forecasts. The latest system, JMA/MRI-CPS3 (Japan Meteorological Agency/Meteorological Research Institute-Coupled Prediction System version 3; "CPS3" in short), was implemented in February 2022 and has been used since for operational forecasts: the "Three-month prediction", "Warm/cold season outlook" and "El Niño outlook". This page outlines the specifications and operational time schedule of CPS3.


CPS3 is an atmosphere/ocean/land/sea ice-coupled prediction system consisting of an initialization system and a forecast model. The atmospheric model is based on GSM2003 (Yonehara et al. 2020), and additional improvements are added for physical parameterization including cumulus convection, cloud and gravity wave. Atmospheric conditions are initialized with JRA-3Q (Kobayashi et al. 2021) for re-forecasts, and Global Analysis (GA) for operation, in which atmospheric conditions are updated with a shorter time delay. CPS3 has a different land-sea mask from these analysis and a unique lake scheme which needs initialization. To avoid initial shock, offline surface simulation for the period to date is separately run and used for forecasts. Ocean and sea ice conditions are initialized with MOVE/MRI.COM-G3, a global ocean 4DVAR analysis (Usui et al. 2015) downscaled to an eddy-permitting resolution (0.25 by 0.25° in longitude and latitude). The ensemble prediction system involves a combination of LAF and initial perturbation as in the previous system (Takaya et al. 2018). Five-member ensemble predictions are made every day from 00UTC, and perturbed atmospheric conditions are determined for each initial time using BGM (Toth and Kalnay 1993, Chikamoto et al. 2007). Ocean perturbations are calculated using 4DVAR minimization history, by which daily analysis error covariance can be approximated (Niwa and Fujii 2020).
Operation Start Date June 2015 February 2022
Atmospheric model Model version GSM1011C GSM2003C
Horizontal Resolution Global TL159 reduced Gaussian grid(~110km) Global TL319 reduced Gaussian grid(~55km)
Vertical levels(model top) 60 levels(0.1hPa) 100 levels(0.01hPa)
Ocean model Model version MRI.COM v3.2 MRI.COM v4.6
Horizontal Resolution 1° (longitude) x 0.3-0.5° (latitude) 0.25° (longitude) x 0.25° (latitude)
Vertical levels 52 levels with a bottom boundary layer 60 levels
Forecast Frequency 13 ensemble members per 5 days 5 ensemble members per day
Initial condition Atmosphere JRA-55 JRA-3Q (hindcasts) and the Global Analysis (GA; forecasts only)
Land Surface JRA-55 Offline model runs forced by JRA-3Q and GA
Ocean MOVE/MRI.COM-G2(3DVAR) MOVE/MRI.COM-G3(Low-res. 4DVAR+High-res. downscaling)
Ensemble generation method Initial Condition perturbation The breeding of growing mode (BGM) method for the atmosphere
The Lagged Average Forecast (LAF) method
BGM for the atmosphere
Ocean perturbations calculated using 4DVAR minimization history
The LAF method
Model Perturbation Stochastic physics parameterization for the atmosphere
Hindcast Period Two initial dates per month for 1979-2014 Two initial dates per month for 1991-2020
Ensemble size Five for each initial date

Real-time operation and ensemble technique

The figure below shows how CPS3 runs in operation. The five-member ensemble is generated once a day from 00UTC analysis. JMA currently uses only three members for 17 LAF dates to make 51-member model statistics products. Registered users can access all members at the download page.

Schematic figure of ensemble configurations of operational suite
Fig. 1 Schematic figure of ensemble configurations of operational suite.

Performance of JMA/MRI-CPS3

To assess forecast performance, a set of hindcasts (re-forecasts) was carried out as in real-time operation (except for a smaller ensemble size of 10 per month). Verification scores are openly published on the hindcast verification page, and registered users can download hindcast data from the download page.


Appendix: Upgrade history of JMA's coupled prediction system

Implemented date Upgrades and Changes
Aug. 1998 The first JMA's CGCM for ENSO prediction implemented to operation.
Jul. 2003 The second CGCM for ENSO prediction (JMA-CGCM02) with an improved atmospheric model and sea-surface flux adjustment.
Jun. 2005 Frequent forecast update (from every 15 days to every 5 days).
Mar. 2008 The Third CGCM for ENSO prediction (JMA/MRI-CGCM1). Oceanic component changed to the new ocean general circulation model "MRI.COM" and data assimilation system "MOVE/MRI.COM-G".
Mar. 2009 Introduced an ocean initial perturbation method. The ensemble size increased from 12 to 30.
Mar. 2010 The first integrated forecast system "JMA/MRI-CPS1" that covers seasonal forecasts in addition to the "El Niño outlook". Introduced an atmospheric perturbation method. The ensemble size increased from 30 to 51.
Jun. 2015 JMA/MRI-CPS2: a dynamical sea ice model, land initialization, historical GHG concentration. The ensemble size per initial date increased from 9 to 13.
Feb. 2022 JMA/MRI-CPS3: 4DVAR ocean and 3DVAR sea ice initialization, an eddy-permitting ocean resolution, improved atmospheric physics, daily forecast update (5 ensemble members per day).

page top

Tokyo Climate Center, Climate Prediction Division.
Copyright(C) 2002. Japan Meteorological Agency. - Legal Notice -