Improving Stream Solute Predictions With a Modified LSTM Model Incorporating Solute Interdependences and Hysteresis Patterns

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 1
• Main Authors: Agrawal, Tarun, Goodwell, Allison, Kumar, Praveen
• Format: Journal Article
• Language: English
• Published: Wiley 01.03.2025
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000383

Surface runoff and infiltrated water en route to the stream interact with dynamic landscape properties, ranging from vegetation and microbial activities to soil and geological attributes. Stream solute concentrations are highly variable and interconnected due to these interactions, flow paths, and residence times, and often exhibit hysteresis with flow. Significant unknowns remain about how point measurements of stream solute chemistry reflect interdependent hydrobiogeochemical and physical processes, and how signatures are encapsulated as nonlinear dynamical relationships between variables. We take a Machine Learning (ML) approach to understand and capture these dynamical relationships and improve predictions of solutes at short and long time scales. We introduce a physical process‐based “flow‐gate” into an Long Short‐Term Memory (LSTM) model, which enables the model to learn hysteresis behaviors if they exist. Further, we use information‐theoretic metrics to detect how solutes are interdependent and iteratively select source solutes that best predict a given target solute concentration. The “flow‐gate LSTM” model improves model predictions (1%–32% decreases in RMSE) relative to the standard LSTM model for all nine solutes included in the study. The predictive improvements from the flow‐gate LSTM model highlight the importance of lagged concentration and discharge relationships for certain solutes. It also indicates a potential limitation in the traditional LSTM model approach since flow rates are always provided as input sources, but this information is not fully utilized. This work provides a starting point for a predictive understanding of geochemical interdependencies using machine‐learning approaches and highlights potential improvements in model architecture. Plain Language Summary Solute concentrations and mass exports from rivers are important to understand and predict, as they impact water quality downstream. Solutes in streamflow include nitrates, calcium, sodium, magnesium, and other cations and anions, and concentrations depend on watershed characteristics such as vegetation, soil, and bedrock composition, flow paths and residence times, microbial activity, water quantity, and human applications of road salts or fertilizers. In other words, solute inputs to rivers vary with time, discharge, and source of water, and solute dynamics are often interrelated due to common sources. This, combined with few high‐resolution observations of stream solutes, makes predictions of solute concentrations challenging. We build on a ML framework, Long Short‐Term Memory (LSTM), to improve solute predictions in two ways. First, we add to the traditional LSTM model architecture to specifically incorporate flow gradient into the model, to better capture hysteresis, where concentrations vary depending on whether flow is increasing or decreasing. Second, to predict a given solute, we detect a set of highly informative solutes, that best predict the target solute without adding redundancies in model inputs. These changes in model input selection and architecture lead to large model improvements, for both long‐term (several weeks) and short‐term (7 hr) predictions. Key Points We propose a modified Long Short‐Term Memory (LSTM) model architecture for stream solute concentration predictions A “flow‐gate” LSTM explicitly accounts for hysteretic behaviors and outperforms the traditional LSTM model We select model inputs with an information theoretic approach to retain maximally informative sources