Author(s): Makoto Nakatsugawa; Riko Sakamoto; Yosuke Kobayashi

Linked Author(s):

Keywords: Dam inflow prediction; Severe flood; Machine learning; Elastic Net; Neural Network

Abstract: We examined machine learning methods to determine which one would be optimal for predicting inflow during severe flood to a dam and reservoir water level. Predictions using machine learning were done for Kanayama Dam, at the upper reaches of the Sorachi River, which is a tributary of the Ishikari River, and for Satsunaigawa Dam, at the upper reaches of the Satsunai River, which is a tributary of the Tokachi River. The predictions were done by using hydrological information for the basins of these two rivers collected at the heavy rainfall disaster of August 2016. As machine learning based prediction methods, RF (Random Forest), FCNN (Fully Connected Neural Network), RNN (Recurrent Neural Network) and regression analysis by Elastic Net, which is a sort of sparse modeling, were examined. As a result, FCNN and Elastic Net demonstrated roughly the same accuracy, with Nash-Sutcliffe (NS) coefficients of 0.7 or greater. With Elastic Net, for cases other than those whose predicted rainfall had indeterminacy, the results were the most accurate and then the NS coefficients were 0.7 or greater. We are sure that obtained results give a promise to improve dam operation for disaster mitigation due to flood.

DOI:

Year: 2020