Deep Learning based Emulator for Rainfall-Runoff-Inundation Model

Abstract

There is an increasing demand for a technological infrastructure that can efficiently design, predict, and control products throughout the product life cycle. In the case of digital twin technology and cyber-physical systems, there is a gap between the operating conditions in the real world and the simulation in the virtual space. In order to reduce this gap, data assimilation, which integrates the real world and virtual space, has a great potential. Data assimilation is mainly used in the field of meteorology and oceanography, and there is a gap in the research of data assimilation based on engineering design, actual operation, and product life cycle. In this study, emulators which reduce the computational cost of physical model-based simulations are used for rainfall-runoff inundation models. The emulator is a simplified model that mimics part or the whole of a physical model based on a process. By building the emulator, it is possible to reduce the computational cost and perform predictions at a high frequency while maintaining simulations based on an elaborate physical model. This study develops an emulator for hydrological model calculations using meteorological data as input, especially for the Rainfall-Runoff-Inundation Model. This study applies deep learning to handle spatio-temporal information and compare it with an existing emulator model using machine learning that was performed using training data for each point.