Physically Informed Machine Learning for Hydrological Modeling Under Climate Nonstationarity

Abstract

There is an understanding in the hydrological sciences community that physical realism is necessary for providing hydrological forecasts under changing conditions (Blöschl et al., 2019; Clark et al., 2016; Milly et al., 2008). At present, however, machine learning (ML) generally provides the best estimates of most hydrological states and fluxes, even in extrapolation (e.g., Best et al., 2015; Kratzert et al., 2019a,b; Nearing et al., 2018). A notable example of this was provided by Kratzert et al. (2019a), who showed that Long Short Term Memory networks (LSTMs) produce, on average, better predictions in basins that did not supply training data (effectively ungaged basins) than a conceptual model well-calibrated to gauge data in individual basins (gaged basins). This is significant in that the 2003-2012 decadal problem of the International Association of Hydrological Sciences (IAHS) was `Prediction in Ungauged Basins' (PUB) (Hrachowitz et al., 2013). Prior to Kratzert et al. (2019a), best practices for PUB required extensive catchment-specific investment (Blöschl, 2016), which is infeasible at large scales (e.g., regional, continental, global).