予測雨量のダウンスケーリングとタンクモデルによる天然ダムの水位予測精度の向上

抄録

In September 2011, Typhoon No.12 (Typhoon Ta las) caused many deep-seated landslides in southern Nara and Wakayama prefectures due to rainfall far exceeding the total precipitation of 1,000 mm, forming landslides dams. Since landslide dam can cause extensive flood damage downstream when overflow erosion occurs, at the Kii Mountain District Sabo Office MLIT, as a risk management, the predicted precipitation and water level have been forecasted using a combination of very short range forecast of precipitation by the JMA, MSM, and GSM, and the storage function method. For the accuracy improvement, this study examined a water level prediction method that combines rainfall prediction by downscaling the MSM guidance and GSM guidance, which are currently the most accurate, with AI and a three-stage tank model that can respond to various hydrological characteristics. This predicted precipitation is based on the analyzed precipitation for the past 10 years with a 1 km grid and 1 hour interval, and a dataset with the same rough timeinterval and spatial resolution as MSM-G and GSM-G. Based on this coarse-resolution guidance rainfall, a convolutional neural network (CNN), a type of deep learning, is used to calculate the predicted precipitation on the same time and spatial scales as the analytical precipitation. The parameters in the three-stage tank model were determined to match the inflow into the landslide dam, which was determined by a water balance analysis based on hydrological observations of the landslide dam. The parameters in the three-stage tank model were determined to match the inflow into the landslide dam as determined by a water balance analysis based on hydrological observations of the landslide dam. Although a tank model can be easily adapted to rainfall with various sizes, the parameter adjustment is complicated. Therefore, the SCE-UA method, one of the optimization methods, was used to obtain the parameters for efficiency. The total precipitation in the AI prediction tends to be underestimated, but the ratio in the AI predcitiction and the observation was close to 1.0, and the three-stage tank model improved the accuracy of both the peak water level difference and the peak time difference in predicting water levels.