STUDY ON PREDICTION OF WATER LEVEL OF LOW FREQUENCY FLOOD ON A LARGE-SCALE RIVER BASED ON THE RANDOM FOREST METHOD

Abstract

 Recently, large-scale floods are increasing in Japan. In Accordingly, we aim to improve the performance of water level prediction to large-scale floods. In this study, we generated water level prediction modelss were using hydrological information on the upper reaches of forecast locations on the Ishikari River and the Tokachi River. At first, we used the Random Forest (RF) method, which is a machine learning method, and we generated water level prediction model were generated with lead times of 6 hours and 12 hours. Next, we selected explanatory variables that are strongly correlated with objective variables were extracted from the RF model. Then, we used this variable to calculate a multiple regression model which named the related factor correlation method. As a result, we could generate a high performance multiple regression model which is Nash-Sutcliffe coefficient score was 0.7 or more on both rivers. Through this study, we have obtained practical and highly accurate findings for forecasts of peak water level and water level rises.