ESTIMATION OF SNOWMELT RUNOFF USING DEEP LEARNING

Abstract

 Prediction of snowmelt water in early spring is very important for water use and disaster prevention managements in heavy snowfall regions. Generally, current snowmelt runoff models have physical processes, and the parameters are calbralted by trial and error due to some uncertainties such as snowfall estimation. Therefore, we estimated the inflow of resourvoir during the snowmelt period by using deep learning for the Sagurigawa dam and the Oishi dam in Niigata prefecuture, which are one of the most heavy snowfall areas in Japan. The inflow were calculated by combining three deep learning model and four input datasets in this study. The reporoducivillties of inflow by deep leraning models were well in normal and heavy snowfall years. However its accuracy decreased in light snow years. Futhermore, it was suggested that the accuracy of the Recurrent Neural Network - based learning model with input data period of 10 days tends to be high, and the air temperature as the input data greatly affected the accuracy.