Post‐Processing the National Water Model with Long Short‐Term Memory Networks for Streamflow Predictions and Model Diagnostics

Abstract

We build three long short‐term memory (LSTM) daily streamflow prediction models (deep learning networks) for 531 basins across the contiguous United States (CONUS), and compare their performance: (1) a LSTM post‐processor trained on the United States National Water Model (NWM) outputs (LSTM_PP), (2) a LSTM post‐processor trained on the NWM outputs and atmospheric forcings (LSTM_PPA), and (3) a LSTM model trained only on atmospheric forcing (LSTM_A). We trained the LSTMs for the period 2004–2014 and evaluated on 1994–2002, and compared several performance metrics to the NWM reanalysis. Overall performance of the three LSTMs is similar, with median NSE scores of 0.73 (LSTM_PP), 0.75 (LSTM_PPA), and 0.74 (LSTM_A), and all three LSTMs outperform the NWM validation scores of 0.62. Additionally, LSTM_A outperforms LSTM_PP and LSTM_PPA in ungauged basins. While LSTM as a post‐processor improves NWM predictions substantially, we achieved comparable performance with the LSTM trained without the NWM outputs (LSTM_A). Finally, we performed a sensitivity analysis to diagnose the land surface component of the NWM as the source of mass bias error and the channel router as a source of simulation timing error. This indicates that the NWM channel routing scheme should be considered a priority for NWM improvement.

Deep learning rainfall–runoff predictions of extreme events

“…These metrics depend heavily on the observed flow characteristics during a particular test period and, because they are less stable, are somewhat less useful in terms of drawing general conclusions. We report them here primarily for continuity with previous studies (Kratzert et al, 2019b(Kratzert et al, , a, 2021Frame et al, 2021;Nearing et al, 2020a; Table 2. Median performance metrics across 498 basins for two separate time split test periods and a test period split by the return period (or probability) of the annual peak flow event (i.e., testing across years with a peak annual event above a 5-year return period, or a 20 % probability of annual exceedance).…”

“…This finding (differences between pure ML and physicsinformed ML) is worth discussing. The project of adding physical constraints to ML is an active area of research across most fields of science and engineering (Karniadakis et al, 2021), including hydrology (e.g., Zhao et al, 2019;Jiang et al, 2020;Frame et al, 2021). It is important to understand that there is only one type of situation in which adding any type of constraint (physically based or otherwise) to a datadriven model can add value: if constraints help optimization.…”

The Data Synergy Effects of Time‐Series Deep Learning Models in Hydrology

“…Recently, deep learning (DL) approaches have proven to be a promising tool in modeling hydrologic dynamics (Shen, 2018; Shen & Lawson, 2021; Sit et al., 2020). Among these, long short‐term memory (LSTM) networks (Hochreiter & Schmidhuber, 1997) present excellent performance in modeling soil moisture (Fang et al., 2017, 2019), streamflow (Feng et al., 2020; Frame et al., 2021; Gauch, Kratzert, et al., 2021; Ha et al., 2021; Kratzert et al., 2019; Nearing, Klotz, et al., 2021; Xiang & Demir, 2020), water table depth (Zhang et al., 2018), water quality variables such as water temperature (Rahmani et al., 2020, 2021) and dissolved oxygen (Zhi et al., 2021), and reservoir modulation (Ouyang et al., 2021). DL can be adapted for tasks like uncertainty quantification (Fang et al., 2020; Li et al., 2021), data assimilation (Fang & Shen, 2020; Feng et al., 2020), and multiscale modeling (Liu et al., 2022).…”

Can machine learning accelerate process understanding and decision‐relevant predictions of river water quality?

“…Many studies in hydrology have attempted to improve predictions of extremes such as floods (Mosavi et al, 2018) using multiple techniques that include using different criteria for model selection (Coulibaly et al, 2001), conditional density estimation networks (Cannon, 2012), training exclusively on extreme events such as historical high-flow data (Fleming et al, 2015), adjustment of ML prediction bias to improve performance on the tails of the distribution (Belitz & Stackelberg, 2021), and using KGML models, for example by training an ML model on simulation data containing extremes that might not exist in the observation data Xie et al, 2021). However, in some cases KGML models may perform worse than traditional DL models; for example, Frame et al (2021) found that an LSTM constrained to conserve mass was not able to predict peak flows as well as the base LSTM, although both models had lower errors than the process-based model used for comparison.…”