Interpretable Machine Learning for Reservoir Water Temperatures in the U.S. Red River Basin of the South

Isabela Suaza-Sierra, Hernan A. Moreno, Luis A De la Fuente, Thomas M. Neeson · 2025

Read Interpretable Machine Learning for Reservoir Water Temperatures in the U.S. Red River Basin of the South by Isabela Suaza-Sierra, Hernan A. Moreno, Luis A De la Fuente, Thomas M. Neeson (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-01837.txt.

Introduction

Isabela Suaza-Sierra, Hernan Moreno, Luis De la Fuente, and Thomas Neeson predict reservoir water temperatures across ten reservoirs. The reservoirs sit in the Red River Basin of the southern United States. The team then insists on understanding the predictions. Temperature governs water management, ecosystem health, and climate resilience, but a black-box forecast alone teaches nothing about the physics.

From more than 10,000 depth-resolved temperature profiles, ensemble and neural models reached high skill, with the best root-mean-square error near 1.2 degrees Celsius. SHAP analysis ranked the drivers: air temperature, depth, wind, and lake volume, in consistent patterns across reservoirs. Kolmogorov-Arnold networks then distilled the behavior into ten progressively richer equations. They are mostly linear and rational in form. Their accuracy rose with complexity, but gained little beyond five predictors. That is a clean simplicity-versus-accuracy trade-off.

Catalog record

CREATORIsabela Suaza-Sierra, Hernan A. Moreno, Luis A De la Fuente, Thomas M. Neeson
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE84 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-01837

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Interpretable Machine Learning for Reservoir Water Temperatures in the U.S. Red River Basin of the South Isabela Suaza-Sierra, Hernan A. Moreno, Luis A De la Fuente, Thomas M. Neeson Abstract Accurate prediction of Reservoir Water Temperature (RWT) is vital for sustainable water management, ecosystem health, and climate resilience. Yet, prediction alone offers limited insight into the governing physical processes. To bridge this gap, we integrated explainable machine learning (ML) with symbolic modeling to uncover the drivers of RWT dynamics across ten reservoirs in the Red River Basin, USA, using over 10,000 depth-resolved temperature profiles. We first employed ensemble and neural models, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP),achieving high predictive skill (best RMSE = 1.20 °C, R² = 0.97). Using SHAP (SHapley Additive exPlanations), we quantified the contribution of physical drivers such as air temperature, depth, wind, and lake volume, revealing consistent patterns across reservoirs. To translate these data-driven insights into compact analytical expressions, we developed Kolmogorov–Arnold Networks (KANs) to symbolically approximate RWT. Ten progressively complex KAN equations were derived, improving from R² = 0.84 using a single predictor (7-day antecedent air temperature) to R² = 0.92 with ten predictors, though gains diminished beyond five, highlighting a balance between simplicity and accuracy. The resulting equations, dominated by linear and rational forms, incrementally captured nonlinear behavior while preserving interpretability. Depth consistently emerged as a secondary but critical predictor, whereas precipitation had limited effect. By coupling predictive accuracy with explanatory power, this framework demonstrates how KANs and explainable ML can transform black-box models into transparent surrogates that advance both prediction and understanding of reservoir thermal dynamics. 1 Introduction Reservoir Water Temperature (RWT) is a fundamental variable in freshwater systems, shaping aquatic organisms’ physiology, behavior, and spatial distribution ( Caissie 2006 ) . It regulates key ecosystem functions including thermal stratification, nutrient cycling, and dissolved oxygen dynamics ( Adrian et al. 2009 ; Davis et al. 2024 ) . Operationally, RWT serves as a critical input for reservoir management, guiding decisions in hydropower generation, water supply, and flood control ( Magee and Wu 2017 ; Imberger and Patterson 1989 ) . Warmer water temperatures can elevate the risk of hypoxia and harmful algal blooms, exacerbating water quality concerns ( Jöhnk et al. 2008 ) . Furthermore, RWT is a sensitive indicator of climate variability and change, responding to changes in atmospheric temperature, wind, and precipitation patterns ( Song et al. 2016 ; Schmid et al. 2014 ) . Modeling RWT thus enhances our capacity to predict and adapt to climate-driven hydrological changes. Water temperature in the epilimnion (the upper, well-mixed layer of a stratified lake) responds rapidly to atmospheric forcing, making it a reliable indicator of short-term hydrometeorological trends ( Robertson and Ragotzkie 1990 ; Piccolroaz et al. 2024 ) . In contrast, hypolimnetic temperatures (those in the cold, dense bottom layer isolated from surface interactions) reflect more complex dynamics shaped by lake morphometry ( Gerten and Adrian 2000 ) , seasonality, and longer-term climate signals ( Robertson and Ragotzkie 1990 ) . The vertical distribution of temperature governs density gradients, which in turn influence vertical mixing, stratification regimes, and thermocline depth ( Woolway et al. 2022 ; Coats et al. 2006 ) . Seasonal stratification regulates nutrient redistribution and biological productivity ( Liu et al. 2018 ) , while physicochemical gradients can directly affect fish health and ecosystem stability ( Jeppesen et al. 2012 ) . Reservoir water temperature is shaped by a complex interplay of meteorological, hydrological, and morphometric drivers, each exerting influence across distinct temporal and spatial scales. Among these, air temperature is consistently identified as the dominant atmospheric driver, controlling surface energy exchanges through sensible heat flux and dictating the net atmospheric heat input to surface waters ( Robertson and Ragotzkie 1990 ; Liu et al. 2018 ) . Diurnal and seasonal variability in air temperature contributes directly to surface water warming rates and the persistence of thermal stratification ( Woolway et al. 2016 ) . Solar radiation, particularly shortwave radiation, penetrates the water column and is absorbed at different depths based on turbidity, colored dissolved organic matter, and particulate content, driving heating in the epilimnion ( Winslow et al. 2017 ; Piccolroaz et al. 2013 ; Woolway and Merchant 2017 ) . In contrast, longwave radiation and evaporative cooling govern nighttime and seasonal cooling cycles, influencing energy loss from the surface and contributing to vertical heat redistribution ( Austin and Allen 2011 ; Stepanenko et al. 2013 ) . Wind speed and direction affect water temperature indirectly but powerfully by driving surface shear stress and vertical turbulent mixing. Wind-induced mixing can deepen the mixed layer and disrupt stratification or cause full overturn, especially in shallower or polymictic systems ( Imberger and Patterson 1989 ; Magee and Wu 2017 ) . In deeper lakes and reservoirs, such mixing is often confined to upper layers, leaving hypolimnetic temperatures less influenced by daily wind dynamics ( Stefan et al. 2001 ) . Hydrological inflows and outflows modulate reservoir temperature profiles through thermal density currents and advective heat transport. Cold snowmelt or warm runoff entering at different times of year can intrude at intermediate depths, creating temporary stratification or thermal barriers within the metalimnion ( Goudsmit et al. 2002 ) . Outflows can remove specific water layers depending on withdrawal elevation, altering temperature distributions and residence time ( Casamitjana et al. 2003 ) . Lastly, reservoir morphometry, including surface area, depth, and hypsographic shape, mediates all these processes by influencing heat storage capacity and stratification stability ( Kirillin and Shatwell 2016 ) . These physical drivers act interactively and often nonlinearly, resulting in complex and site-specific thermal dynamics. Their influence varies not only between reservoirs but also with time, particularly under changing climate conditions and reservoir regulation schemes. Understanding their mechanisms is essential for accurately modeling water temperature, anticipating ecological consequences, and informing management decisions in the context of warming and hydrological change ( Wang et al. 2023 ; Woolway et al. 2021 ) . Traditional RWT models, both physics-based and empirical, simulate complex interactions involving heat fluxes (radiation, evaporation, conduction), mass exchanges (inflows, outflows, vapor), and mechanical forces such as wind-driven mixing ( Woolway et al. 2021 ) . These models often simulate vertical mixing and stratification influenced by temperature, chemistry, and biology ( Tanentzap et al. 2007 ) , but suffer from limited scalability, site dependency, and heavy computational demands ( Munz and Schmidt 2017 ) . Empirical models offer computational efficiency but typically oversimplify the system, failing to capture nonstationary and nonlinear behavior ( Rashid and Beecham 2019 ) . Even widely used models like the General Lake Model (GLM) require extensive calibration and are sensitive to input uncertainty under climate scenarios ( Fenocchi et al. 2018 ; Abbasian et al. 2020 ) . ML approaches, including Random Forest (RF) ( Breiman 2001 ) , Extreme Gradient Boosting (XGBoost) ( Chen and Guestrin 2016 ) , and Multilayer Perceptron (MLP) ( Rumelhart et al. 1986 ) have gained traction for their ability to learn from large datasets and generalize across hydrological systems. Examples include RF applications in evaporation forecasting from reservoirs in Ethiopia ( Eshetu et al. 2023 ) and permeability prediction in tight

Read the complete text