9:15 AM - 9:30 AM
[HDS09-02] Impact of the number of ensemble members on the accuracy of flood risk forecasting
Keywords:flood risk, ensemble forecasts, flood forecasting
Floods caused by Quasi-stationary convective bands (QSCB) have caused significant damage. Flood risk information needs to be available about 12 hours in advance, especially late at night and dawn. The Japan Meteorological Agency provides flood risk information up to 6 hours in advance with the operational flood forecast system "Kiki-kuru". However, providing accurate deterministic weather forecasts for 12 hours is difficult. Therefore, ensemble weather forecasting is necessary to increase lead times. The JMA provide the 21-member operational weather forecast (MEPS). However, it is still difficult to forecast QSCB with the MEPS. Some researchers reported that a large ensemble of weather forecasts was useful for QSCB forecasts. However, only a few studies investigated whether such large ensemble forecasts could predict floods caused by QSCB. Even minor errors in predicting the locations and amount of rainfall in weather forecasts can lead to significant errors in flood forecasting.
This study investigated the accuracy of flood forecasting using a large ensemble of weather forecasts for the flood in the Kuma River on 4 July 2020. Target rives were 36 tributaries connected to the Kuma River. The flood forecasting model was the "Kiki-kuru", and the precipitation dataset was 100, 1000 members ensemble forecasts and MEPS.
We classified the tributaries by the size of the basin: 12 rivers were less than 20 km2, 17 rivers were less than 40 km2, six rivers were less than 100 km2, and only 553 km2 of the Kawabe River was more than 100 km2.
In rivers of 20 km2 or less, irrespective of the damage of the flood, the 1000-member experiment (1000MEM) showed a trend of increased risk of flooding during the peak flooding period calculated using the Radar-AMeDAS system. Meanwhile, the 100-member experiment (100MEM) showed an increased risk of flooding several hours later than the peak period by the Radar-AMeDAS system. The experiment with MEPS (21MEM) could not predict the timing of floods. The 1000MEM showed a higher probability of floods exceeding the warning level than the 100MEM experiment; for 21MEM, floods exceeding the warning were only a few rivers. A similar trend was seen for rivers larger than 40 km2. On the Kawaba River, 1000MEM predicted a 60% probability of the warning level's flood occurring at the peak 12 hours in advance. Meanwhile, 100MEM predicted a 40% probability of a warning-level flood two hours later than the peak. 21MEM failed to predict the warning level's flood.
This study showed that the 1000MEM was better than the 100MEM. The 21MEM showed the potential to predict whether flooding would occur 12 hours ahead without being able to predict the peak time. The results suggest that the impact of the large ensemble on flood forecasting needs to be validated by increasing the sample of case studies and catchment areas.
This study investigated the accuracy of flood forecasting using a large ensemble of weather forecasts for the flood in the Kuma River on 4 July 2020. Target rives were 36 tributaries connected to the Kuma River. The flood forecasting model was the "Kiki-kuru", and the precipitation dataset was 100, 1000 members ensemble forecasts and MEPS.
We classified the tributaries by the size of the basin: 12 rivers were less than 20 km2, 17 rivers were less than 40 km2, six rivers were less than 100 km2, and only 553 km2 of the Kawabe River was more than 100 km2.
In rivers of 20 km2 or less, irrespective of the damage of the flood, the 1000-member experiment (1000MEM) showed a trend of increased risk of flooding during the peak flooding period calculated using the Radar-AMeDAS system. Meanwhile, the 100-member experiment (100MEM) showed an increased risk of flooding several hours later than the peak period by the Radar-AMeDAS system. The experiment with MEPS (21MEM) could not predict the timing of floods. The 1000MEM showed a higher probability of floods exceeding the warning level than the 100MEM experiment; for 21MEM, floods exceeding the warning were only a few rivers. A similar trend was seen for rivers larger than 40 km2. On the Kawaba River, 1000MEM predicted a 60% probability of the warning level's flood occurring at the peak 12 hours in advance. Meanwhile, 100MEM predicted a 40% probability of a warning-level flood two hours later than the peak. 21MEM failed to predict the warning level's flood.
This study showed that the 1000MEM was better than the 100MEM. The 21MEM showed the potential to predict whether flooding would occur 12 hours ahead without being able to predict the peak time. The results suggest that the impact of the large ensemble on flood forecasting needs to be validated by increasing the sample of case studies and catchment areas.