5:15 PM - 7:15 PM
[AAS10-P03] Study on the Utilization of Flow-Dependent Background Error Covariance Using Meso-ensemble Prediction System
Keywords:Ensemble forecast using outputs from MEPSMEPS
Improving the accuracy of heavy rainfall forecasts, such as senjo-kousuitai, is an urgent issue. Various new observation techniques and data assimilation methods are being developed to enhance accuracy. At the National Research Institute for Earth Science and Disaster Resilience (NIED), a three-dimensional variational method (3DVAR) with relatively low computational cost has been used to generate initial conditions in real-time, enabling high-frequency short-term rainfall predictions every 10 minutes.
In 3DVAR, climatological background error covariance derived from several months of statistical processing is utilized. In contrast, recent advancements in ensemble-based assimilation methods allow the use of flow-dependent background error covariance, which enables anisotropic data assimilation while considering inter-variable correlations. Although ensemble-based assimilation methods offer significant advantages, the computational cost of ensemble forecasting remains high.
Yokota et al. (2024, JSMJ) proposed effectively utilizing operational meso-ensemble prediction systems (MEPS) to reduce the computational cost of ensemble forecasting while developing a flow-dependent 3DVAR, demonstrating the crucial role of MEPS in improving the accuracy of heavy rainfall predictions. However, external institutions outside the Japan Meteorological Agency (JMA) must rely on MEPS provided by the Meteorological Business Support Center (MBSC), where basic variables (wind, pressure, temperature, and humidity) are thinned to reduce data size, making direct use of MEPS as initial conditions for ensemble prediction difficult.
In this study, we developed a method to achieve a 21-member ensemble prediction by combining forecasts from the mesoscale model (MSM) with the Meso-Ensemble Prediction System (MEPS) provided by the Meteorological Business Support Center (MBSC). The procedure is as follows:
Calculate the ensemble mean and perturbations of the 21 MEPS members. At this stage, relative humidity is converted to water vapor mixing ratio. Use MSM as the ensemble mean and apply the perturbations obtained in step (1) to generate 21 members (Member 1 is approximately the same as MSM but not exactly identical). If the water vapor mixing ratio becomes negative after perturbation, it is set to 0 kg/kg. Convert temperature and water vapor mixing ratio to relative humidity for each member. If saturation occurs, the water vapor mixing ratio is adjusted to the saturation mixing ratio. Apply steps (1) and (2) to surface data, specifically horizontal wind speed and temperature. Since MEPS includes forecasts at 3-hour intervals, repeat steps (1)–(4) for the forecast data and use them as lateral boundary conditions for CReSS simulations.
For a linear rainband event that occurred in Kumamoto Prefecture on July 3, 2023, we conducted a 3-hour forecast using MEPS at 00 UTC. The initial analysis was performed with a horizontal resolution of 3 km, covering the entire Kyushu region (312 km × 360 km × 50 layers, model top = 20.6 km). Figure 1 (left) shows the 3-hour accumulated rainfall from XRAIN (provided by the Ministry of Land, Infrastructure, Transport, and Tourism), while Figure 1 (right) shows the ensemble mean of the 3-hour accumulated rainfall. The area with rainfall exceeding 50 mm was generally predicted correctly, but CReSS predicted the location of heavy rainfall exceeding 100 mm slightly to the north.
Figure 2 displays the spread of various variables at 03 UTC, with the top row showing CReSS and the bottom row showing MEPS. From left to right, the variables presented are the east-west wind component, north-south wind component, potential temperature (temperature in MEPS), and water vapor mixing ratio. The spread in CReSS is generally more localized than in MEPS, and it is concentrated around areas where heavy rainfall was predicted. Since the resolutions of MEPS and CReSS differ, an exact match is not expected. However, understanding these characteristic differences is the objective of this study for future development of flow-dependent 3DVAR.
In the future, we plan to compare the assimilation increments obtained using LETKF with pseudo-observation data at 03 UTC and those obtained using En-3DVAR.
In 3DVAR, climatological background error covariance derived from several months of statistical processing is utilized. In contrast, recent advancements in ensemble-based assimilation methods allow the use of flow-dependent background error covariance, which enables anisotropic data assimilation while considering inter-variable correlations. Although ensemble-based assimilation methods offer significant advantages, the computational cost of ensemble forecasting remains high.
Yokota et al. (2024, JSMJ) proposed effectively utilizing operational meso-ensemble prediction systems (MEPS) to reduce the computational cost of ensemble forecasting while developing a flow-dependent 3DVAR, demonstrating the crucial role of MEPS in improving the accuracy of heavy rainfall predictions. However, external institutions outside the Japan Meteorological Agency (JMA) must rely on MEPS provided by the Meteorological Business Support Center (MBSC), where basic variables (wind, pressure, temperature, and humidity) are thinned to reduce data size, making direct use of MEPS as initial conditions for ensemble prediction difficult.
In this study, we developed a method to achieve a 21-member ensemble prediction by combining forecasts from the mesoscale model (MSM) with the Meso-Ensemble Prediction System (MEPS) provided by the Meteorological Business Support Center (MBSC). The procedure is as follows:
Calculate the ensemble mean and perturbations of the 21 MEPS members. At this stage, relative humidity is converted to water vapor mixing ratio. Use MSM as the ensemble mean and apply the perturbations obtained in step (1) to generate 21 members (Member 1 is approximately the same as MSM but not exactly identical). If the water vapor mixing ratio becomes negative after perturbation, it is set to 0 kg/kg. Convert temperature and water vapor mixing ratio to relative humidity for each member. If saturation occurs, the water vapor mixing ratio is adjusted to the saturation mixing ratio. Apply steps (1) and (2) to surface data, specifically horizontal wind speed and temperature. Since MEPS includes forecasts at 3-hour intervals, repeat steps (1)–(4) for the forecast data and use them as lateral boundary conditions for CReSS simulations.
For a linear rainband event that occurred in Kumamoto Prefecture on July 3, 2023, we conducted a 3-hour forecast using MEPS at 00 UTC. The initial analysis was performed with a horizontal resolution of 3 km, covering the entire Kyushu region (312 km × 360 km × 50 layers, model top = 20.6 km). Figure 1 (left) shows the 3-hour accumulated rainfall from XRAIN (provided by the Ministry of Land, Infrastructure, Transport, and Tourism), while Figure 1 (right) shows the ensemble mean of the 3-hour accumulated rainfall. The area with rainfall exceeding 50 mm was generally predicted correctly, but CReSS predicted the location of heavy rainfall exceeding 100 mm slightly to the north.
Figure 2 displays the spread of various variables at 03 UTC, with the top row showing CReSS and the bottom row showing MEPS. From left to right, the variables presented are the east-west wind component, north-south wind component, potential temperature (temperature in MEPS), and water vapor mixing ratio. The spread in CReSS is generally more localized than in MEPS, and it is concentrated around areas where heavy rainfall was predicted. Since the resolutions of MEPS and CReSS differ, an exact match is not expected. However, understanding these characteristic differences is the objective of this study for future development of flow-dependent 3DVAR.
In the future, we plan to compare the assimilation increments obtained using LETKF with pseudo-observation data at 03 UTC and those obtained using En-3DVAR.