Routing flood waves through the river network utilizing physics-guided machine learning and the Muskingum-Cunge Method

Abstract:

Rainfall-runoff models have shown the ability to generate subbasin runoff with high accuracy. However, the subbasin-based long short-term memory (LSTM) models does not scale properly to large rivers because of two reasons: (i) the heterogeneity of inputs (forcing and attributes) is too great that a model with uniform inputs no longer makes sense; (ii) the in-channel routing process is difficult to model with rainfall-runoff LSTM models. The Muskingum-Cunge method uses a river network, which we now call river graph, to model in-channel routing without the need for a neural network. This paper routes observed runoff through the channel network using the Muskingum-Cunge method with empirically-estimated parameters. We share deeper insights of the river network using this physics-guided neural network. We show promising results from the coupled model that can now deal with large river basins. Opportunities exist for our model to produce scalable flood routing results.