EMULATION EVALUATION OF URBAN RUNOFF MODEL BY DEEP LEARNING FOR THE VIRTUAL HYDROGRAPH WITH OBSERVATION NOISE

Abstract

 Due to climate change associated with global warming, short-term heavy rainfall is increasing, and there are great expectations for the application of deep learning models to flood forecasting, which are being applied to various fields. Therefore, in this paper, it was decided to confirm how accurately the runoff model could be emulated by the deep learning model when the observation data contained observation noise. Virtual rainfall was used as input data, and virtual runoff height including observation noise with clear characteristics was used as teacher data. As the learning data, we evaluated the emulation performance of the runoff model by the deep learning model when the number of learning floods and the number of nodes in the hidden layer were changed. When the number of learning floods is small, we learned to match the noise component. Moreover, it was found that increasing the number of learning floods is effective to grasp the characteristics of the hydrograph, which is the true value without observation noise.