When a typhoon appears, it is important for coastal disaster prevention to conduct an accurate storm surge forecast (SSF). However, it is difficult to closely observe meteorological fields in time and space on the open ocean, and the initial values used for typhoon forecast might greatly cause the difference between the typhoon forecast and the actual one. Since differences in meteorological fields cause differences in storm surges, the accurate SSF is still a great challenge.

Therefore, in this study, we used the Regional Oceanic Modeling System (ROMS) to develop a storm surge numerical model which successively assimilates ocean observation data. In this model, the calculation is revised every time after observed values are obtained, and so is expected to suppress differences in prediction from occurring. The assimilation method uses the Ensemble Kalman Filter (EnKF), which derives the analytical value from multiple first guesses with different initial conditions.

For model verification, we conducted the following twin experiment for the case of Typhoon Lan in 2017. 1) We calculated the storm surge with the meteorological analytic value MSM-GPV (grid interval: 5 km) as an external force (calculation period: between 0:00, October 21, 2017 and 12:00, October 24, 2017; calculation area: lat. 22.5° to 47.5° N, long. 125° to 149° E; grid interval: 5 km), handled the results as true values, and then made the observed values by adding Gaussian noises on the true values. 2) We used the LAF method to set seven different initial conditions. With each condition as a starting point, we calculated the storm surge with the meteorological forecast value GSM-GPV (grid interval: 20 km) as an external force. 3) Regarding the calculations, we assimilated the observed values every hour at points where GPS-mounted wave buoys are positioned.

As a result, we confirmed that the assimilation affected the overall calculation area, even though the observed values were given at the points. However, the assimilation did not improve forecast accuracy.

In the future, we will investigate the following optimal model setting: applying the Breeding method for making new initial conditions, increasing the number of ensemble members, and shortening the assimilation interval.