DEPENDENCY ON NETWORK STRUCTURE AND INFORMATION DENSITY OF DEEP LEARNING BASED RIVER STAGE PREDICTION

Abstract

 A few of recent studies have applied deep learning technique for flood forecast. This study utilizes Recurrent Neural Network (RNN) with Gated Recurrent Units (GRUs) to hindcast the Kanto-Tohoku Flood in September 2015. Additionally, based on linear reservoir function, it applies Exponential Filtering (EF) as a preprocessor of input data to transform the statistical characteristics of input variable (i.e., rainfall) to of the target variable (i.e., river water stage). Compared with Feed Forward Network (FFN) model and RNN model without EF, the proposed model outperforms the flood events in Kinu river basin. In particular, it results reduced error for the highest water level which is 3.43 meter higher than of the highest level during the training period . Also, we investigate dependency of prediction skill on neural network structure and input data information density in Tone-river and Teshio-river basin, which shows critical number to predictability of water level and rainfall observation sites differs among target stations and basins.