Improve Streamflow Prediction Research Articles

Streamflow Prediction at the Intersection of Physics and Machine Learning: A Case Study of Two Mediterranean‐Climate Watersheds

AbstractAccurate streamflow predictions are essential for water resources management. Recent studies have examined the use of hybrid models that integrate machine learning models with process‐based (PB) hydrologic models to improve streamflow predictions. Yet, there are many open questions regarding optimal hybrid model construction, especially in Mediterranean‐climate watersheds that experience pronounced wet and dry seasons. In this study, we performed model benchmarking to (a) compare hybrid model performance to PB and machine learning models and (b) examine the sensitivity of hybrid model performance to PB model parameter calibration, structural complexity, and variable selection. Hybrid models were generated by post‐processing process‐based models using Long Short‐Term Memory neural networks. Models were benchmarked within two northern California watersheds that are managed for both municipal water supplies and aquatic habitat. Though model performance varied substantially by watershed and error metric, calibrated hybrid models frequently outperformed both the machine learning model (for 72% of watershed‐model‐metric combinations) and the calibrated process‐based models (for 79% of combinations). Furthermore, hybrid models were relatively insensitive to PB model calibration and structural complexity, but sensitive to PB model variable selection. Our results demonstrate that hybrid models can improve streamflow prediction in Mediterranean‐climate watersheds. Additionally, hybrid model insensitivity to PB model parameter calibration and structural complexity suggests that uncalibrated or less complex PB models could be used in hybrid models without any loss of streamflow prediction accuracy, improving model construction efficiency. Moreover, hybrid model sensitivity to the selection of PB model variables suggests a strategy for diagnosing poorly performing PB model components.

Streamflow prediction in ungauged catchments through use of catchment classification and deep learning

Streamflow prediction in ungauged catchments is a challenging task in hydrological studies. Recently, data-driven models have demonstrated their superiority over traditional hydrological models in predicting streamflow in ungauged catchments. However, previous studies have overlooked the similarities between the training and the target catchments. Therefore, this study explores the role of catchment similarity in regionalization modeling using the publicly available CAMELS dataset. We employed the dynamic time warping-based KMeans (DTW-KMeans) time-series clustering technique to cluster the streamflow data from gauged catchments. We utilized the long short-term memory (LSTM) neural network to construct regional models for different classes of gauged catchment. Additionally, the mapping relationship between gauged catchment classes and static attributes was established using the random forest (RF). By combining the trained RF model with the static attributes of an ungauged catchment, we determined its class and used the corresponding regional LSTM to predict streamflow. To evaluate the effectiveness of the framework, we applied the classification-based regionalization modeling (CRM) and non-classification-based regionalization modeling (NRM) approach for comparison. The results indicate that: (1) The DTW-KMeans-based catchment classification method is generally accurate and reasonable; (2) the complexity of the LSTM model and the number of training catchments should be appropriately matched to improve streamflow prediction; and (3) catchment similarity plays a crucial role in regionalization modeling, the proportion of training catchments with high similarity to ungauged catchments significantly affects prediction results.

Temporal Fusion Transformers for streamflow Prediction: Value of combining attention with recurrence

Over the past few decades, the hydrology community has witnessed notable advancements in streamflow prediction, particularly with the introduction of cutting-edge machine-learning algorithms. Recurrent neural networks, especially Long Short-Term Memory (LSTM) networks, have become popular due to their capacity to create precise forecasts and realistically mimic the system dynamics. Attention-based models, such as Transformers, can learn from the entire data sequence concurrently, a feature that LSTM does not have. This work tests the hypothesis that combining recurrence with attention can improve streamflow prediction. We set up the Temporal Fusion Transformer (TFT) architecture, a model that combines both of these aspects and has never been applied in hydrology before. We compare the performance of LSTM, Transformers, and TFT over 2,610 globally distributed catchments from the recently available Caravan dataset. Our results demonstrate that TFT indeed exceeds the performance benchmark set by the LSTM and Transformers for streamflow prediction, when models are trained individually for each basin. Additionally, being an explainable AI method, TFT helps in gaining insights into the streamflow generation processes.

Hydrologic applicability of satellite-based precipitation estimates for irrigation water management in the data-scarce region

Reliable precipitation estimates are crucial for planning and managing water resources, monitoring hydrologic extremes, and fulfilling irrigation water requirements. Accurate precipitation estimates are particularly challenging in complex mountain terrains, where monitoring gauges are often sparsely distributed due to their remote locations, and high installation and long-term operation costs. Recent advances in satellite-based precipitation estimates offer promising opportunities to improve our understanding of hydrologic processes and their applications for irrigation water management. Several datasets are available varying considerably in terms of their data sources, quality control methods, estimation procedure, and spatiotemporal resolutions. Choosing the most suitable dataset for a particular application is a complex task. In this study, we (1) evaluate the performance of six satellite-based precipitation estimates (SPEs): i) CHIRPS v2.0, ii) CMORPH v1.0, iii) ERA5, iv) IMERG v6, v) MSWEP v2.8, and vi) PERSIANN-CDR against the gauge precipitation using continuous statistical and categorical indices, (2) integrate SPEs with a calibrated semi-distributed hydrologic model to predict streamflow, and (3) demonstrate practical implications of improved streamflow prediction for irrigation water management in the central Himalayan region, Nepal. Our results illustrate that satellite-based precipitation estimates have competitive performance in capturing a wide range of rainfall characteristics, with demonstrated variability across river basins and time scales. There are no significant discrepancies observed in satellite-based precipitation estimates for estimating irrigation water requirements for the three major crops (maize, wheat, and paddy) during the cropping period across the selected river basins, showing a greater promise for irrigation water management planning and decision making.

Exploring the potential of deep learning for streamflow forecasting: A comparative study with hydrological models for seasonal and perennial rivers

Improving streamflow prediction plays a significant role in flood warning, mitigation and development purposes. Therefore, this paper aims to compare the prediction capability of a useful conceptual hydrologic model hydrologiska byråns vattenbalansavdelning (HBV) enhanced and equipped with an automatic calibration procedure (Genetic Algorithm) along with two well-known deep learning (DL) algorithms, including Long-Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). This study was concentrated on the daily streamflow of two river basins with diverse climatic and hydrological conditions, namely the Halilrood River with fragmented and intermittent flows (located in Iran) and the Tweed River with a permanent stream flow (Located across the border of Scotland and northern England). These were particularly selected to address both seasonal and perennial riverine cases. For reproducing the streamflow, 10-year records of the daily discharge, precipitation, and temperature (from satellite and ground resources), and evapotranspiration were used as the main input data. The accuracy of the implemented models was assessed using several indices. The findings show that 1) the LSTM and GRU show better accuracy than the employed hydrologic model for estimating daily discharges (especially in the seasonal river basin); 2) the DL approaches need only meteorological data inputs (precipitation and temperature) and are independent of involved uncertainties in hydrological modeling such as physical characteristics of basins; and 3) best streamflow simulations were achieved by the LSTM (best agreement with the observations). In general, LSTM has the highest priority to be connected with online hydrological rainfall-runoff simulations and dynamical forecast models.

On optimization of calibrations of a distributed hydrological model with spatially distributed information on snow

Abstract. In northern cold-temperate countries, a large portion of annual streamflow is produced by spring snowmelt, which often triggers floods. It is important to have spatial information about snow variables such as snow water equivalent (SWE), which can be incorporated into hydrological models, making them more efficient tools for improved decision-making. The present research implements a unique spatial pattern metric in a multi-objective framework for calibration of hydrological models and attempts to determine whether raw SNODAS (SNOw Data Assimilation System) data can be utilized for hydrological model calibration. The spatial efficiency (SPAEF) metric is explored for spatially calibrating SWE. Different calibration experiments are performed combining Nash–Sutcliffe efficiency (NSE) for streamflow and root-mean-square error (RMSE) and SPAEF for SWE, using the Dynamically Dimensioned Search (DDS) and Pareto Archived Dynamically Dimensioned Search multi-objective optimization (PADDS) algorithms. Results of the study demonstrate that multi-objective calibration outperforms sequential calibration in terms of model performance (SWE and discharge simulations). Traditional model calibration involving only streamflow produced slightly higher NSE values; however, the spatial distribution of SWE could not be adequately maintained. This study indicates that utilizing SPAEF for spatial calibration of snow parameters improved streamflow prediction compared to the conventional practice of using RMSE for calibration. SPAEF is further implied to be a more effective metric than RMSE for both sequential and multi-objective calibration. During validation, the calibration experiment incorporating multi-objective SPAEF exhibits enhanced performance in terms of NSE and Kling–Gupta efficiency (KGE) compared to calibration experiment solely based on NSE. This observation supports the notion that incorporating SPAEF computed on raw SNODAS data within the calibration framework results in a more robust hydrological model. The novelty of this study is the implementation of SPAEF with respect to spatially distributed SWE for calibrating a distributed hydrological model.

Assessing effect of best management practices in unmonitored watersheds using the coupled SWAT-BiLSTM approach

In order to enhance the simulation of BMPs (Best Management Practices) reduction effects in unmonitored watersheds, in this study, we combined the physically-based hydrological model Soil & Water Assessment Tool (SWAT) and the data-driven model Bi-directional Long Short-Term Memory (Bi-LSTM), using the very-high-resolution (VHR) Land Use and Land Cover (LULC) dataset SinoLC-1 as data input, to evaluate the feasibility of constructing a water environment model for the Ba-River Basin (BRB) in central China and improving streamflow prediction performance. In the SWAT-BiLSTM model, we calibrated the top five SWAT parameters sorted by P-Value, allowing SWAT to act as a transfer function to convert meteorological data into base flow and storm flow, serving as the data input for the Bi-LSTM model. This optimization improved the Bi-LSTM's learning process for the relationship between the target and explanatory variables. The daily streamflow prediction results showed that the hybrid model had 9 regions rated as "Very good," 2 as "Good," 2 as "Satisfactory," and 1 as "Unsatisfactory" among the 14 regions. The model achieved an NSE of 0.86, R2 of 0.85, and PBIAS of −2.71% for the overall daily streamflow prediction performance during the verification period of the BRB. This indicates that the hybrid model has high predictive accuracy and no significant systematic bias, providing a sound hydrodynamic environment for water quality simulation. The simulation results of different BMPs scenarios showed that in the scenarios with only one BMP measure, stubble mulch had the best reduction effect, with average reductions of 17.83% for TN and 36.17% for TP. In the scenarios with a combination of multiple BMP measures, the combination of stubble mulch, soil testing and formula fertilization, and vegetative filter strip performed the best, achieving average reductions of 42.71% for TN and 50.40% for TP. The hybrid model provides a novel approach to simulate BMPs' reduction effects in regions without measured hydrological data and has the potential for wide application in BMP-related decision-making.

Uncertainty quantification of machine learning models to improve streamflow prediction under changing climate and environmental conditions

Machine learning (ML) models, and Long Short-Term Memory (LSTM) networks in particular, have demonstrated remarkable performance in streamflow prediction and are increasingly being used by the hydrological research community. However, most of these applications do not include uncertainty quantification (UQ). ML models are data driven and can suffer from large extrapolation errors when applied to changing climate/environmental conditions. UQ is required to quantify the influence of data noises on model predictions and avoid overconfident projections in extrapolation. In this work, we integrate a novel UQ method, called PI3NN, with LSTM networks for streamflow prediction. PI3NN calculates Prediction Intervals by training 3 Neural Networks. It can precisely quantify the predictive uncertainty caused by the data noise and identify out-of-distribution (OOD) data in a non-stationary condition to avoid overconfident predictions. We apply the PI3NN-LSTM method in the snow-dominant East River Watershed in the western US and in the rain-driven Walker Branch Watershed in the southeastern US. Results indicate that for the prediction data which have similar features as the training data, PI3NN precisely quantifies the predictive uncertainty with the desired confidence level; and for the OOD data where the LSTM network fails to make accurate predictions, PI3NN produces a reasonably large uncertainty indicating that the results are not trustworthy and should avoid overconfidence. PI3NN is computationally efficient, robust in performance, and generalizable to various network structures and data with no distributional assumptions. It can be broadly applied in ML-based hydrological simulations for credible prediction.

A multi-criteria approach for improving streamflow prediction in a rapidly urbanizing data scarce catchment

ABSTRACT This study advocates a multi-criteria approach to improve the streamflow predictions in a data-scarce catchment of Chennai metropolitan city of India using the Soil Water and Assessment Tool (SWAT). The remotely sensed evapotranspiration (ET) data, groundwater recharge estimation, and parameter regionalization were used to improve model prediction. Dynamic change of Land Use and Land Cover (LULC) was accounted for along with multi-parameter calibration for reducing the uncertainty in model parameters. The results revealed an improved streamflow prediction accuracy by 10%, especially in the prediction of medium and high flows with the Nash-Sutcliffe efficiency of 0.60. The enhanced parameters were regionalized to ungauged sub-basins and validated using a measured flow event downstream of regionalization with 15% prediction uncertainty. This semi-arid catchment is dominated by ET (58%) and runoff (27%) in the region's hydrology. The finding of this study can be applied to improve the hydrological modelling and predictions in data-scarce regions.

Coupling ANFIS with ant colony optimization (ACO) algorithm for 1-, 2-, and 3-days ahead forecasting of daily streamflow, a case study in Poland

Finding an efficient and reliable streamflow forecasting model has always been an important challenge for managers and planners of freshwater resources. The current study, based on an adaptive neuro-fuzzy inference system (ANFIS) model, was designed to predict the Warta river (Poland) streamflow for 1 day, 2 days, and 3 days ahead for a data set from the period of 1993–2013. The ANFIS was additionally combined with the ant colony optimization (ACO) algorithm and employed as a meta-heuristic ANFIS-ACO model, which is a novelty in streamflow prediction studies. The investigations showed that on a daily scale, precipitation had a very weak and insignificant effect on the river’s flow variation, so it was not considered as a predictor input. The predictor inputs were selected by the autocorrelation function from among the daily streamflow time lags for all stations. The predictions were evaluated with the actual streamflow data, using such criteria as root mean square error (RMSE), normalized RMSE (NRMSE), and R2. According to the NRMSE values, which ranged between 0.016–0.006, 0.030–0.013, and 0.038–0.020 for the 1-day, 2-day, and 3-day lead times, respectively, all predictions were classified as excellent in terms of accuracy (prediction quality). The best RMSE value was 1.551 m3/s and the highest R2 value was equal to 0.998, forecast for 1-day lead time. The combination of ANFIS with the ACO algorithm enabled to significantly improve streamflow prediction. The use of this coupling can averagely increase the prediction accuracies of ANFIS by 12.1%, 12.91%, and 13.66%, for 1-day, 2-day, and 3-day lead times, respectively. The current satisfactory results suggest that the employed hybrid approach could be successfully applied for daily streamflow prediction in other catchment areas.

Representing Bidirectional Hydraulic Continuum Between the Stream and Hillslope in the National Water Model for Improved Streamflow Prediction

AbstractAlthough hydraulic groundwater (GW) theory has been recognized as a promising tool for understanding the role of the aquifer(s) in the surface‐subsurface hydrologic cycle, the integrated modeling community still lacks a proper hydrologic structure to apply the well‐studied theory to large‐scale hydrologic predictions. This study aims to present a novel hydrologic structure that enables the Boussinesq equation‐based depiction of the bidirectional stream‐hillslope processes for applying hydraulic GW theory to large‐scale model configurations. We integrated the BE3S's (Hong et al., 2020, https://doi.org/10.1029/2020wr027571) representation scheme of the catchment‐scale stream‐hillslope continuum into the National Water Model (NWM) and applied the modified NWM (i.e., the NWM‐BE3S) to three major basins in Texas (i.e., the Trinity, Brazos, and Colorado River basins). Since the NWM currently relies on a single reservoir model for baseflow simulation, we used the Boussinesq aquifer as an alternative subsurface hydrology routine and evaluated its predictive skill and efficacy. We identified that the implemented Boussinesq aquifer(s) in the NWM‐BE3S yielded noticeable improvements in predicting streamflow for aquifers that exhibited higher nonlinearities in the observed recessions. The varying degree of improvements in streamflow predictions per the recession nonlinearities demonstrated not only (a) the algorithmic enhancement of subsurface hydrology (physics) but also (b) the applicability of the Boussinesq theory‐based depiction of the stream‐hillslope two‐way continuum. We diagnosed each stream's state based on the bidirectional stream‐hillslope exchanges and identified the dominant processes (i.e., river infiltration or baseflow) that were represented spatially in the NWM‐BE3S.

Incorporating IMERG satellite precipitation uncertainty into seasonal and peak streamflow predictions using the Hillslope Link hydrological model

In global applications and data sparse regions, which comprise most of the earth, hydrologic model-based flood monitoring relies on precipitation data from satellite multisensor precipitation products or numerical weather forecasts. However, these products often exhibit substantial errors during the meteorological conditions that lead to flooding, including extreme rainfall. The propagation of precipitation forcing errors to predicted runoff and streamflow is scale-dependent and requires an understanding of the autocorrelation structure of precipitation errors, since error autocorrelation impacts the accumulation of precipitation errors over space and time in hydrologic models. Previous efforts to account for satellite precipitation uncertainty in hydrologic models have demonstrated the potential for improving streamflow estimates; however, these efforts use satellite precipitation error models that rely heavily on ground reference data such as rain gages or weather radar and do not characterize the nonstationarity of precipitation error autocorrelation structures. This work evaluates a new method, the Space-Time Rainfall Error and Autocorrelation Model (STREAM), which stochastically generates possible true precipitation fields, as input to the Hillslope Link Model to generate ensemble streamflow estimates. Unlike previous error models, STREAM represents the nonstationary and anisotropic autocorrelation structure of satellite precipitation error and does not use any ground reference to do so. Ensemble streamflow predictions are compared with streamflow generated using satellite precipitation fields as well as a radar-gage precipitation dataset during peak flow events. Results demonstrate that this approach to accounting for precipitation uncertainty effectively characterizes the uncertainty in streamflow estimates and reduces the error of predicted streamflow. Streamflow ensembles forced by STREAM improve streamflow prediction nearly to the level obtained using ground-reference forcing data across basin sizes.

Improvement of springtime streamflow prediction using a snow hydrology model aided with USDA SNOTEL and in-situ snowpit observations

AbstractEstimating the streamflow driven by snowmelt in rugged mountain watersheds is difficult. Challenges are associated with the limited observations of hydrologic and meteorological datasets and inadequate implementation of the snow hydrology models. This study aims to improve streamflow prediction during the snowmelt season using a snow hydrology model aided by field observations. When the point-based weather forcing data and in-situ snowpit measurements exist in or near a small-scale (2–3 km2) watershed, the hydrologic model demonstrated an improved streamflow prediction during the snowmelt period. A snow hydrology model was applied to the Senator Beck Basin (SBB) in Colorado to improve the streamflow prediction. A temperature index method was implemented in the hydrological model to accommodate the snowmelt routine, which releases water as a multiplication factor for a grid temperature surplus above the melting point. The temperature index was adjusted using in-situ snowpit observations collected in the SBB by the NASA SnowEx Year-1 campaign in February 2017. Using the determined temperature index and weather forcing data from the nearby USDA snow observation telemetry station, the Nash-Sutcliffe Efficiency of the simulated streamflow was elucidated with a value of 0.88 against the observed streamflow during April 1–22, 2017.

WITHDRAWN: Incorporating IMERG satellite precipitation uncertainty into seasonal and peak streamflow predictions using the Hillslope Link hydrological model

In global applications and data sparse regions, which comprise most of the earth, hydrologic model-based flood monitoring relies on precipitation data from satellite multisensor precipitation products or numerical weather forecasts. However, these products often exhibit substantial errors during the meteorological conditions that lead to flooding, including extreme rainfall. The propagation of precipitation forcing errors to predicted runoff and streamflow is scale-dependent and requires an understanding of the autocorrelation structure of precipitation errors, since error autocorrelation impacts the accumulation of precipitation errors over space and time in hydrologic models. Previous efforts to account for satellite precipitation uncertainty in hydrologic models have demonstrated the potential for improving streamflow estimates; however, these efforts use satellite precipitation error models that rely heavily on ground reference data such as rain gages or weather radar and do not characterize the nonstationarity of precipitation error autocorrelation structures. This work evaluates a new method, the Space-Time Rainfall Error and Autocorrelation Model (STREAM), which stochastically generates possible true precipitation fields, as input to the Hillslope Link Model to generate ensemble streamflow estimates. Unlike previous error models, STREAM represents the nonstationary and anisotropic autocorrelation structure of satellite precipitation error and does not use any ground reference to do so. Ensemble streamflow predictions are compared with streamflow generated using satellite precipitation fields as well as a radar-gage precipitation dataset during peak flow events. Results demonstrate that this approach to accounting for precipitation uncertainty effectively characterizes the uncertainty in streamflow estimates and reduces the error of predicted streamflow. Streamflow ensembles forced by STREAM improve streamflow prediction nearly to the level obtained using ground-reference forcing data across basin sizes.

A Model Combination Approach for Improving Streamflow Prediction

A Model Combination Approach for Improving Streamflow Prediction

Streamflow Prediction in Highly Regulated, Transboundary Watersheds Using Multi‐Basin Modeling and Remote Sensing Imagery

Despite the potential of remote sensing for monitoring reservoir operation, few studies have investigated the extent to which reservoir releases can be inferred across different spatial and temporal scales. Through evaluating 21 reservoirs in the highly regulated Greater Mekong region, remote sensing imagery was found to be useful in estimating daily storage volumes for within‐year and over‐year reservoirs (correlation coefficients [CC] ≥ 0.9, normalized root mean squared error [NRMSE] ≤ 31%), but not for run‐of‐river reservoirs (CC < 0.4, 40% ≤ NRMSE ≤ 270%). Given a large gap in the number of reservoirs between global and local databases, the proposed framework can improve representation of existing reservoirs in the global reservoir database and thus human impacts in hydrological models. Adopting an Integrated Reservoir Operation Scheme within a multi‐basin model was found to overcome the limitations of remote sensing and improve streamflow prediction at ungauged cascade reservoir systems where previous modeling approaches were unsuccessful. As a result, daily regulated streamflow was predicted competently across all types of reservoirs (median values of CC = 0.65, NRMSE = 8%, and Kling‐Gupta efficiency [KGE] = 0.55) and downstream hydrological stations (median values of CC = 0.94, NRMSE = 8%, and KGE = 0.81). The findings are valuable for helping to understand the impacts of reservoirs and dams on streamflow and for developing more useful adaptation measures to extreme events in data sparse river basins.

Improving streamflow prediction in the WRF-Hydro model with LSTM networks

Researchers have attempted to use machine learning algorithms to replace physically based models for streamflow prediction. Although existing studies have contributed to improving machine learning methods, they still have weaknesses, such as large dataset requirements and overfitting. Therefore, we propose an approach that combines the Weather Research and Forecasting hydrological modeling system (WRF-Hydro) and the Long Short-Term Memory (LSTM) network, i.e., WRF-Hydro-LSTM, to improve streamflow simulations. In this approach, LSTM was employed to predict the residual errors of WRF-Hydro; in contrast, the conventional approach with LSTM predicts streamflow directly. Here, we performed numerical experiments to predict the inflow of Soyangho Lake in South Korea using WRF-Hydro-LSTM, WRF-Hydro-only, and LSTM-only. WRF-Hydro-LSTM and LSTM-only showed better results (NSE = 0.95 and R greater than 0.96) compared to WRF-Hydro-only (NSE = 0.72 and R = 0.88); however, in terms of the percent bias, WRF-Hydro-LSTM had a better value (1.75) than LSTM-only (17.36). While the LSTM-only follows objective functions and not physical principles, WRF-Hydro-LSTM simulates residual errors and efficiently decreases uncertainties that are inherent with conventional methods. Furthermore, a sensitivity test on the training dataset indicated that the correlation coefficient and NSE value were not overly sensitive, but the PBIAS value differed substantially depending on the training set. This study demonstrates that WRF-Hydro-LSTM is particularly useful for representing real-world physical constraints and thus can potentially improve streamflow prediction compared to using either of the two approaches exclusively.

Random forests-based error-correction of streamflow from a large-scale hydrological model: Using model state variables to estimate error terms

To improve streamflow predictions, researchers have implemented updating procedures that correct predictions from a simulation model using machine learning methods, in which simulated streamflow and meteorological data are used as predictors. Few studies however have included an extensive set of meteorological and hydrological state variables simulated by the simulation model. We developed and evaluated a Random Forests (RF)-based approach to correct predictions from a global hydrological model PCR-GLOBWB. From PCR-GLOBWB, meteorological input as well as its simulated hydrological state variables were used as predictors in the RF to estimate errors of PCR-GLOBWB streamflow predictions, which were then applied to correct simulated hydrograph. The RF was trained and applied separately at three streamflow gauging stations in the Rhine basin with different physiographic characteristics. Daily streamflow simulations from an uncalibrated PCR-GLOBWB run were improved by applying the RF-based error-correction model (KGE improved from 0.37 to 0.62 to 0.76–0.89, NSE from 0.19 to 0.39 to 0.64–0.80). A similar improvement was found in the simulations from a calibrated PCR-GLOBWB run (KGE 0.72–0.87 and NSE 0.60–0.78). The PCR-GLOBWB state variables that are informative to the improvement differed between catchments. Variables related to groundwater are informative in catchments dominated by the sedimentary basins characterizing large aquifers, while snow cover and surface water state variables are informative in a nival regime with large lakes. Here we quantified the improvement from combining a process-based and machine learning approach.

Groundwater‐Mediated Memory of Past Climate Controls Water Yield in Snowmelt‐Dominated Catchments

AbstractAccelerating warming, changes in the amount, timing, and form of precipitation, and rapidly growing populations highlight the need for improved predictions of snowmelt‐driven water supplies. Although decadal‐scale trends in reduced streamflow are common, minimal progress has been made in improving streamflow prediction on the annual time scales on which management decisions are made. Efficient allocation of dwindling supplies requires incorporating rapidly evolving knowledge of streamflow generation into parsimonious models capable of improving prediction on seasonal, annual, and multiyear time scales of water resource management. We address this need using long‐term streamflow and climate records in 12 catchments averaging 90 years of observations and totaling more than 1,080 site‐years of data. These catchments experience similar regional climate forcing each year, but are diverse enough to represent broad ranges in precipitation, temperature, vegetation, and geology characteristic of much of the western US. We find that January baseflow across all catchments exhibits a coherent, quasi‐decadal periodicity that presumably is indicative of groundwater response to decadal climate. Although the direct contribution of this discharge to streamflow is small, interannual variability in groundwater discharge is a consistently strong predictor of runoff efficiency suggesting that antecedent groundwater storage alters precipitation routing to streamflow. Incorporating antecedent groundwater storage with precipitation and melt dynamics in multiple linear regression models reduces uncertainty in annual runoff from approximately 40% to &lt;5%. These simple models, using readily available data, provide immediately useful tools for water managers to anticipate and respond to streamflow variability on time scales of 1 to 10 years.

Untangling hybrid hydrological models with explainable artificial intelligence

AbstractHydrological models are valuable tools for developing streamflow predictions in unmonitored catchments to increase our understanding of hydrological processes. A recent effort has been made in the development of hybrid (conceptual/machine learning) models that can preserve some of the hydrological processes represented by conceptual models and can improve streamflow predictions. However, these studies have not explored how the data-driven component of hybrid models resolved runoff routing. In this study, explainable artificial intelligence (XAI) techniques are used to turn a ‘black-box’ model into a ‘glass box’ model. The hybrid models reduced the root-mean-square error of the simulated streamflow values by approximately 27, 50, and 24% for stations 17120000, 27380000, and 33680000, respectively, relative to the traditional method. XAI techniques helped unveil the importance of accounting for soil moisture in hydrological models. Differing from purely data-driven hydrological models, the inclusion of the production storage in the proposed hybrid model, which is responsible for estimating the water balance, reduced the short- and long-term dependencies of input variables for streamflow prediction. In addition, soil moisture controlled water percolation, which was the main predictor of streamflow. This finding is because soil moisture controls the underlying mechanisms of groundwater flow into river streams.