CREATION AND VERIFICATION OF A PRETRAINED MODEL FOR RIVER FLOOD PREDICTIONS

Abstract

 This study demonstrates that two-type pretrained models built with artificial neural network (ANN) for reasonable predictions of unprecedented floods in major rivers in Kyushu region, Japan that has similar hydrological features using a small amount of flood datasets. One pretrained model, created by the flood datasets observed in dam-affected rivers, provided better predictions than those of the convention ANN for short lead times and small amounts of the flood datasets. Another pretrained model was created by all flood datasets, available in the major rivers and even no-dam rivers. The difference between the pretrained models shows that the latter model performed better predictions than the former model doing only in the limited condition such as longer lead time and relatively large amount of the datasets.