Integrating Flow and Solute Flux Dynamics in an Adaptive LSTM Model for Stream Chemistry Predictions

Abstract

High‐frequency measurements of stream solute dynamics reveal fluctuations across a multitude of time‐scales. At the storm‐event time‐scales they expose solute specific rapid mobilization or dilution effects, in addition to hysteresis. For various applications, predictions of these dynamics at measurement sites are desired to inform impending events such as harmful algal blooms. To aid in such efforts, we present a machine learning adaptive model. It builds on the well‐established Long Short‐Term Memory (LSTM) model, but introduces important modifications to capture dynamical attributes such as hysteresis, mobilization, and dilution. These modifications incorporate flow and flux gates that embed flow‐ and flux‐gradients directly into the traditional model architecture alongside the input, output, and forget gates. These gates are activated when flow or flux gradients or both exceed certain thresholds. The success of these activations in improving model performance indicates that they capture processes that better inform predictability of dynamical attributes embedded in the observed solute variability. The study is performed using data from a RiverLab, a stream chemistry laboratory located on the bank of a stream, at two sites. The first site is the Upper Sangamon Basin in Illinois, U.S.A., and the second in Orgeval, France. Both are tile drained loess covered agricultural watersheds, but of different drainage areas and in different climates. We assess the model performance using various metrics aligned with our goal to capture key dynamical features. The results justify the need for going beyond the standard formulations of LSTM to better accommodate the nuanced dynamics of natural systems.