Physics informed neural networks for 1D flood routing

Abstract

Machine learning methods have been widely and successfully applied in hydrological problems. Most of the methods, such as artificial neural networks, have been focused on estimating hydrological data based on observation over time. Even though these models provide good results, it can be observed that results become unreliable when the training dataset is small or when input data is significantly out of range compared to the training data. Therefore, a new approach is presented, in which artificial neural networks are trained to satisfy physical laws. This is conducted by a novel method called physics-informed neural networks (PINNs), in which physical principles are embedded in a custom loss function. This paper presents the application of physics informed neural networks for solving 1D flood wave propagation in open channels. The research has shown promising results.