**3.5.4 WWRP Beijing Olympics 2008 RDP project**

An international research project of the World Weather Research Programme (WWRP), Beijing 2008 Olympics Research and Development Project (B08RDP; Duan et al. 2012) was conducted in conjunction with the Beijing 2008 Olympic Games. The main component of B08RDP was short-range forecasting of up to 36 h based on mesoscale EPSs with a horizontal resolution of 15 km. Six institutions from five countries [MRI, NCEP, Meteorological Service of Canada (MSC), Central Institute for Meteorology and Geodynamics (ZAMG) of Austria, National Meteorological Center (NMC) of CMA, and Chinese Academy of Meteorological Sciences (CAMS)] participated in the project, and were requested to run their EPSs for a forecast time of up to 36 h, starting every day at 1200 UTC.

Prior to the 2008 intercomparison period (one month, from 25 July to 23 August 2008), MRI developed five initial perturbation methods: (1) a downscaling of JMA's operational oneweek EPS (WEP; Saito et al., 2010a), (2) a targeted global model singular vector (GSV; Yamaguchi et al., 2009; Hara, 2010) method, (3) MSV method based on the adjoint model of NHM (Kunii, 2010a), (4) a mesoscale breeding of growing modes (MBD) method based on the NHM forecast (Saito, 2007), and (5) a local ensemble transform Kalman filter using NHM (LETKF; Miyoshi and Aranami, 2006; Seko, 2010). Saito et al. (2011a) objectively compared the results of the ensemble forecasts made with these five methods by evaluating the evolution of the ensemble spreads, the RMSE of the ensemble mean. GSV was selected as the initial perturbation method, considering its performance for weak to moderate rain QPF and the RMSE characteristics. The initial condition of the MRI/JMA system was prepared by applying Meso 4DVAR to the B08RDP area (Kunii et al., 2010a).

Kunii et al. (2011) reported the results of the international EPS intercomparison. Verification was performed using the MEP outputs interpolated into a common verification domain. For all systems, the ensemble spreads grew as the forecast time increased, and the ensemble mean improved the RMSEs compared with individual control forecasts in the verification against the analysis fields. MRI/JMA's EPS and the control run had the best performance among the six EPS systems for predicting surface conditions (2m temperature and relative humidity) and weak to moderate rains, in terms of RSMEs against the initial condition and the threat scores.

Details of MRI and JMA's activities in B08FDP/RDP have been published as an MRI Technical Report (Saito et al., 2010b).

### **3.5.5 Ensemble Kalman filter**

The ensemble Kalman filter (EnKF) technique is a new method of data assimilation that employs ensemble prediction to estimate forecast error. Miyoshi and Aranami (2006) applied four-dimensional expansion of LETKF to NHM and performed data assimilation experiments in a perfect model scenario with 5-km grid spacing. The analysis equation for LETKF is:

$$\mathbf{X}^{a} = \overline{\mathbf{x}}^{f}\boldsymbol{e} + \delta \mathbf{X}^{f} \overline{\mathbf{P}}^{a} (\mathbf{H}\delta \mathbf{X})^{T} \mathbf{R}^{-1} (\mathbf{y}^{o} - \overline{H(\mathbf{x}^{f})}) \boldsymbol{e} + \delta \mathbf{X}^{f} \mathbf{T}. \tag{19}$$

Here, the overbar means the ensemble mean, and **X**f the ensemble perturbation matrix. **H** is the tangent linear of the observation operator, and *e* is an *m*-dimensional row vector. **P**a with tilde is the analysis error covariance matrix in the space spanned by forecast ensemble perturbations obtained by eigenvalue decomposition.

This system, NHM-LETKF, was modified by Seko (2010) and was tested as the initial perturbation generator for the mesoscale ensemble prediction in the B08RDP project.

Saito et al. (2012) examined the effects of LBPs on the MBD method and LETKF for mesoscale ensemble prediction. Introducing LBPs in the data assimilation cycles of LETKF improved the ensemble spread, the ensemble mean accuracy, and the performance of precipitation forecast. The accuracy of the LETKF analyses was compared with those of the Meso 4D-VAR analyses. With LBPs in the LETKF cycles, the RMSEs of the forecasts from the LETKF analyses improved, and some became comparable to those of the Meso 4D-VAR analyses.

Seko et al. (2011) performed data assimilation experiments with NHM-LETKF for an intense local rainfall event near the city of Kobe, western Japan, on 28 July 2008. They assimilated GEONET GPS TPWV data with conventional observation data. Adding GPS TPWV data tended to increase low-level water vapor and improve the precipitation forecast. The experiment with 5-km resolution generated a rainfall band in western Japan that was not reproduced using conventional data alone. The experiment with 1.6-km resolution effectively reproduced the observed band of intense rainfall.
