Streamflow Forecasting Research Articles

Daily Runoff Forecasting Using Novel Optimized Machine Learning Methods

Accurate runoff forecasting is crucial for effective water resource management, yet existing models often face challenges due to the complexity of hydrological systems. This study addresses these challenges by introducing a novel bio-inspired metaheuristic algorithm, Artificial Rabbits Optimization (ARO), integrated with various machine learning (ML) models for runoff forecasting in the Carson and Chehalis River basins. The ARO-ML models are rigorously evaluated, demonstrating their superior ability to enhance forecasting accuracy compared to widely used ML techniques, including gradient boosting (GB), multi-layer perceptron (MLP), and K-nearest neighbor regression (KNN). In the Carson River, the GB model achieves the highest forecasting accuracy, which is significantly improved by ARO, resulting in a 24.8% reduction in root mean square error (RMSE). The MLP model also benefits notably from ARO, with RMSE improvements of 4.8% and a substantial 48.9% reduction in mean absolute error (MAE). The ARO-optimized KNN model shows exceptional performance, surpassing the baseline KNN during testing. In the Chehalis River Basin, ARO integration leads to substantial improvements in both GB and MLP models across various performance metrics during training and testing. Crucially, the results indicate that in both river basins, rainfall in preceding days has the most significant influence on streamflow forecasts, followed by air temperature. This study advances hydrological forecasting by validating the efficacy of ARO-ML models, offering a robust framework for enhanced predictive accuracy and providing deeper insights into the key drivers of streamflow in diverse hydrological contexts.

Investigating the Performance of the Informer Model for Streamflow Forecasting

Recent studies have shown the potential of transformer-based neural networks in increasing prediction capacity. However, classical transformers present several problems such as computational time complexity and high memory requirements, which make Long Sequence Time-Series Forecasting (LSTF) challenging. The contribution to the prediction of time series of flood events using deep learning techniques is examined, with a particular focus on evaluating the performance of the Informer model (a particular implementation of transformer architecture), which attempts to address the previous issues. The predictive capabilities of the Informer model are explored and compared to statistical methods, stochastic models and traditional deep neural networks. The accuracy, efficiency as well as the limits of the approaches are demonstrated via numerical benchmarks relating to real river streamflow applications. Using daily flow data from the River Test in England as the main case study, we conduct a rigorous evaluation of the Informer efficacy in capturing the complex temporal dependencies inherent in streamflow time series. The analysis is extended to encompass diverse time series datasets from various locations (>100) in the United Kingdom, providing insights into the generalizability of the Informer. The results highlight the superiority of the Informer model over established forecasting methods, especially regarding the LSTF problem. For a forecast horizon of 168 days, the Informer model achieves an NSE of 0.8 and maintains a MAPE below 10%, while the second-best model (LSTM) only achieves −0.63 and 25%, respectively. Furthermore, it is observed that the dependence structure of time series, as expressed by the climacogram, affects the performance of the Informer network.

Daily Streamflow Forecasting Using AutoML and Remote-Sensing-Estimated Rainfall Datasets in the Amazon Biomes

Daily Streamflow Forecasting Using AutoML and Remote-Sensing-Estimated Rainfall Datasets in the Amazon Biomes

Bridging Hydrological Ensemble Simulation and Learning Using Deep Neural Operators

AbstractEnsemble‐based simulation and learning (ESnL) has long been used in hydrology for parameter inference, but computational demands of process‐based ESnL can be quite high. To address this issue, we propose a deep neural operator learning approach. Neural operators are generic machine learning algorithms that can learn functional mappings between infinite‐dimensional spaces, providing a highly flexible tool for scientific machine learning. Our approach is built upon DeepONet, a specific deep neural operator, and is designed to address several common problems in hydrology, namely, model parameter estimation, prediction at ungaged locations, and uncertainty quantification. Here we demonstrate the effectiveness of our DeepONet‐based workflow using an existing large model ensemble created for an eastern U.S. watershed that is instrumented with 10 streamflow gages. Results suggest DeepONet achieves high efficiency in learning an ML surrogate model from the model ensemble, with the modified Kling‐Gupta Efficiency exceeding 0.9 on holdout test sets. Parameter inference, carried out using the trained DeepONet surrogate model and genetic algorithm, also yields robust results. Additionally, we formulate and train a separate DeepONet model for physics‐informed, seq‐to‐seq streamflow forecasting, which further reduces biases in the pre‐trained DeepONet surrogate model. While this study focuses primarily on a single watershed, our approach is general and may be extended to enable learning from model ensembles across multiple basins or models. Thus, this research represents a significant contribution to the application of hybrid machine learning in hydrology.

Advanced streamflow forecasting for Central European Rivers: The Cutting-Edge Kolmogorov-Arnold networks compared to Transformers

Accurate streamflow forecasting is crucial for effective water resource management, flood mitigation, and maintaining ecological balance, especially in Central Europe’s major rivers. This study investigates the performance of two advanced deep learning algorithms—Kolmogorov-Arnold Network (KAN) and Transformers—using long-term hydrological data from four key rivers: the Rhine, Danube, Elbe, and Oder. These rivers, characterized by varying flow regimes and significant human and climatic impacts, serve as vital arteries for both commerce and ecosystems.The dataset comprises historical daily streamflow data, alongside derived input parameters such as moving averages and flow rate changes, used to predict short-term (1 to 7 days) river discharges. The KAN model, leveraging learnable spline-based activation functions, was developed to enhance accuracy and interpretability in capturing complex hydrological patterns. In contrast, the Transformer model uses advanced attention mechanisms, which excel in handling sequential data with long-range dependencies.Results show that the KAN model significantly outperforms the Transformer in short-term forecasts (up to 3 days). For 1-day forecasts, the KAN achieved R2 values of 0.975 for the Rhine, 0.956 for the Danube, 0.992 for the Elbe, and 0.969 for the Oder, with MAPE values ranging from 3.20 % to 8.09 %. At the 3-day horizon, the KAN’s R2 values remained high, demonstrating its robustness. However, as the forecasting horizon extends to 7 days, the performance of both models converges.These findings underscore the methodological novelty of KAN, which provides enhanced short-term predictive capabilities compared to existing methods, offering valuable insights for improving water resource management strategies in complex hydrological settings.

Hybrid Multivariate Machine Learning Models for Streamflow Forecasting: A Two-Stage Decomposition–Reconstruction Framework

Hybrid Multivariate Machine Learning Models for Streamflow Forecasting: A Two-Stage Decomposition–Reconstruction Framework

Improving streamflow forecasting in semi-arid basins by combining data segmentation and attention-based deep learning

The increasing threats of flash floods and water scarcity in semi-arid regions necessitate high-quality streamflow forecasting with process-based or data-driven models. However, temporal heterogeneity of hydrological patterns and infrequent flood events present challenges for accurate streamflow forecasting. To address this issue, we purpose an effective modeling approach that combines a novel data segmentation method, BPX, with an attention-based deep learning (DL) model. BPX integrates Bai-Perron analysis and XGBoost to accurately identify dry and wet periods from hydro-meteorological variables both in the past and future times. The DL model introduces an attention mechanism to temporal convolutional network (TCN) to enhance its ability in handling heterogeneous time-series data. We demonstrate the effectiveness of our approach using nearly 20 years of hydro-meteorological data from the Wei River basin in China. Various modeling methods are compared, including two process-based models (Xinanjiang and GR4J) and four DL models (classical RNN, GRU, LSTM and TCN) with or without the attention module. After systematic benchmarking, it is found that both the BPX segmentation and the attention mechanism can improve streamflow forecasting with DL. Among the various modeling approaches, BPX-TCN exhibits the best overall performance throughout the prediction period, achieving the highest Kling-Gupta efficiency of 0.91 (compared to 0.45–0.88 for the other models), while BPX-TCN-attention provides more accurate predictions of floods and multi-step predictions. In similar applications in semi-arid regions, our approach may serve as a valuable reference where temporal heterogeneity is significant and poses challenges to traditional modeling methods.

Data‐driven artificial intelligence‐based streamflow forecasting, a review of methods, applications, and tools

AbstractData‐driven artificial intelligence (DDAI) prediction has gained much attention, especially in recent years, because of its power and flexibility compared to traditional approaches. In hydrology, streamflow forecasting is one of the areas that took advantage of utilizing DDAI‐based forecasting, given the weakness of the old approaches (e.g., physical‐based approaches). Since many different techniques and tools have been used for streamflow forecasting, there is a new way to explore them. This manuscript reviews the recent (2011–2023) applications of DDAI in streamflow prediction. It provides a background of DDAI‐based techniques, including machine learning algorithms and methods for pre‐processing the data and optimizing or enhancing the machine learning approaches. We also explore the applications of DDAI techniques in streamflow forecasting. Finally, the most common tools for utilizing DDAI techniques in streamflow forecasting are presented.

A hybrid SARIMA-Prophet model for predicting historical streamflow time-series of the Sobat River in South Sudan

Accurate river streamflow forecasting is pivotal for effective water resource planning, infrastructure design, utilization, optimization, and flood planning and warning. Streamflow prediction remains a difficult task due to several factors such as climate change, topography, and lack of observed data in some cases. This paper investigates and evaluates the individual performances of the seasonal auto-regressive integrated moving average (SARIMA) and Prophet models in forecasting the streamflow of the Sobat River and proposes a hybrid SARIMA-Prophet model to leverage the strengths of both approaches. Using the augmented Dickey-Fuller (ADF) and the Kwiatkowski-Phillips-Schmidt-Shin (KPSS) tests, the flow of the Sobat River was found to be stationary. The performance of the models was then assessed based on their residual errors and predictive accuracy using the mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R2). Residual analysis and prediction capabilities revealed that Prophet slightly edged SARIMA in terms of prediction efficacy; however, both models struggled to effectively capture extreme values, resulting in significant overestimations and slight underestimations. The hybrid SARIMA-Prophet model significantly reduced residual variability, achieving a lower MAE of 4.047 m3/s, RMSE of 6.17 m3/s, and a higher R2 of 0.92 than did the SARIMA (MAE: 5.39 m3/s, RMSE: 8.70 m3/s, R2: 0.85) and Prophet (MAE: 5.35 m3/s, RMSE: 8.32 m3/s, and R2: 0.86) models. This indicates that the hybrid model handles both long-term patterns and short-term fluctuations more effectively than the individual models. The findings of the present study highlight the potential of hybrid SARIMA-Prophet models for streamflow forecasting in terms of accuracy and reliability, thus contributing to more effective water resource management and planning, particularly in the Sobat River.

Application of weather post-processing methods for operational ensemble hydrological forecasting on multiple catchments in Canada

Hydrological forecasts contain biases that need to be addressed for their effective use in operational decision-making in water resources management. Performing post-processing allows reducing the overall systematic bias while improving the distribution and accuracy of hydrological forecasts. In this study, a Quantile Mapping (QM) post-processing method was applied on weather forecasts following three temporal configurations (monthly, seasonal, and annual) of the quantile mapping scheme. The evaluation encompasses 20 catchments in southern Canada, employing a leave-one-out approach with the QM method on ECMWF ensemble weather forecasts spanning 2015–2020 inclusively. These processed forecasts are subsequently utilized as forcings for eight hydrological models, generating ensemble streamflow forecasts over a 6-year period with a lead time of 10 days and a sub-daily timestep of 6 h. The performance of the QM method is mainly assessed using the Continuous Ranked Probability Score (CRPS) metric, in complement with a forecast reliability score (ABDU) and a forecast sharpness metric (NMIQR). Significant improvements are discerned in precipitation forecasts upon the application of QM. Notably, these improvements are translated into enhanced hydrological forecasts for over half of the catchments studied (55 %). Surprisingly, no discernible differences in performance are observed among the three QM configurations in most catchments. Interestingly, there are watersheds where the implementation of QM exhibit either poorer or no change in performance and sharpness compared to raw forecasts.

Performance of data-driven models based on seasonal-trend decomposition for streamflow forecasting in different climate regions of Türkiye

This study examines the ability of different methods such as machine learning, ensemble models, and metaheuristic algorithms to predict streamflow. For this purpose, five different methods were used: Artificial Neural Networks (ANN), Support Vector Machines (SVM), Adaptive Boosting, Particle Swarm Optimization (PSO), and BSPSO hybridized with Band Similarity (BS), a relatively new method. Additionally, the impact of seasonality and trend components obtained through Seasonal-Trend decomposition using LOESS (locally weighted regression and scatterplot smoothing) (STL) data decomposition technique on prediction success was investigated. Models were developed in three basins with three different climate characteristics: continental, temperate, and arid. The results showed higher prediction success in input structures including seasonality and trend components. While higher prediction successes were achieved at Karasu in the continental climate class and Körkün in the temperate climate class, model performances were lower at Küçük Muhsine in the arid climate class. While the most successful modeling for Küçük Muhsine (NSE = 0.696) and Karasu stations (NSE = 0.811) was obtained with the BSPSO method, the SVM method produced the best results for Körkün station (NSE = 0.818). Moreover, BSPSO models outperformed the prediction successes obtained by using PSO alone for each scenario at all three stations. The percentages of the BSPSO method improving the prediction success according to the NSE metric ranged from 3.53% to 17.40% at Küçük Muhsine, 0.49%–3.72% at Karasu, and 1.24%–7.24% at Körkün. The competitive results achieved by the BSPSO approach compared to ANN and SVM in flow prediction constitute the innovative aspect of this study.

Ensemble learning of decomposition-based machine learning models for multistep-ahead daily streamflow forecasting in northwest China

ABSTRACT Accurate daily streamflow forecasts remain challenging in arid regions. A Bayesian model averaging (BMA) ensemble learning strategy was proposed to forecast 1-, 2-, and 3-day-ahead streamflow in Dunhuang Oasis, northwest China. The efficiency of BMA was compared with four decomposition-based machine learning and deep learning models. Satisfactory forecasts were achieved with all proposed models at all lead times; however, based on Nash-Sutcliffe efficiency values of 0.976, 0.967, and 0.957, BMA achieved the greatest accuracy for 1-, 2-, and 3-day-ahead streamflow forecasts, respectively. Uncertainty analysis confirmed the reliability of BMA in yielding consistently accurate streamflow forecasts. Thus, BMA could provide an efficient alternative approach to multistep-ahead daily streamflow forecasting. The incorporation of data decomposition techniques (e.g. variational mode decomposition) and deep learning algorithms (e.g. deep belief network) into BMA may provide worthy technical references for supervised learning of streamflow systems in data-scarce regions.

Improving Hurricane Intensity and Streamflow Forecasts in Coupled Hydrometeorological Simulations by Analyzing Precipitation and Boundary Layer Schemes

Abstract Hurricanes have been the most destructive and expensive hydrometeorological event in U.S. history, causing catastrophic winds and floods. Hurricane dynamics can significantly impact the amount and spatial extent of storm precipitation. However, the complex interactions of hurricane intensity and precipitation and the impacts of improving hurricane dynamics on streamflow forecasts are not well established yet. This paper addresses these gaps by comprehensively characterizing the role of vertical diffusion in improving hurricane intensity and streamflow forecasts under different planetary boundary layer, microphysics, and cumulus parameterizations. To this end, the Weather Research and Forecasting (WRF) atmospheric model is coupled with the WRF hydrological (WRF-Hydro) model to simulate four major hurricanes landfalling in three hurricane-prone regions in the United States. First, a stepwise calibration is carried out in WRF-Hydro, which remarkably reduces streamflow forecast errors compared to the U.S. Geological Survey (USGS) gauges. Then, 60 coupled hydrometeorological simulations were conducted to evaluate the performance of current weather parameterizations. All schemes were shown to underestimate the observed intensity of the considered major hurricanes since their diffusion is overdissipative for hurricane flow simulations. By reducing the vertical diffusion, hurricane intensity forecasts were improved by ∼39.5% on average compared to the default models. These intensified hurricanes generated more intense and localized precipitation forcing. This enhancement in intensity led to ∼16% and ∼34% improvements in hurricane streamflow bias and correlation forecasts, respectively. The research underscores the role of improved hurricane dynamics in enhancing flood predictions and provides new insights into the impacts of vertical diffusion on hurricane intensity and streamflow forecasts. Significance Statement Despite significant recent improvements, numerical weather prediction models struggle to accurately forecast hurricane intensity and track due to many reasons such as inaccurate physical parameterization for hurricane flows. Furthermore, the performance of existing physics schemes is not well studied for hurricane flood forecasting. This study bridges these knowledge gaps by extensively evaluating different physical parameterizations for hurricane track, intensity, and flood forecasts using an atmospheric model coupled with a hydrological model. Then, a reduced diffusion boundary layer scheme is developed, making remarkable improvements in hurricane intensity forecasts due to the overdissipative nature of the considered schemes for major hurricane simulations. This reduced diffusion model is shown to significantly enhance hurricane flood forecasts, indicating the significance of hurricane dynamics on its induced precipitation.

Comparative Analysis of Snowmelt-Driven Streamflow Forecasting Using Machine Learning Techniques

The rapid advancement of machine learning techniques has led to their widespread application in various domains, including water resources. However, snowmelt modeling remains an area that has not been extensively explored. In this study, we propose a state-of-the-art (SOTA) deep learning sequential model, leveraging a Temporal Convolutional Network (TCN), for snowmelt forecasting of the Hindu Kush Himalayan (HKH) region. To evaluate the performance of our proposed model, we conducted a comparative analysis with other popular models, including Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and Transformer models. Furthermore, nested cross-validation (CV) was used with five outer folds and three inner folds, and hyperparameter tuning was performed on the inner folds. To evaluate the performance of the model, the Mean Absolute Error (MAE), Root-Mean-Square Error (RMSE), R square (R2), Kling–Gupta Efficiency (KGE), and Nash–Sutcliffe Efficiency (NSE) were computed for each outer fold. The average metrics revealed that the TCN outperformed the other models, with an average MAE of 0.011, RMSE of 0.023, R2 of 0.991, KGE of 0.992, and NSE of 0.991 for one-day forecasts of streamflow. The findings of this study demonstrate the effectiveness of the proposed deep learning model as compared to traditional machine learning approaches for snowmelt-driven streamflow forecasting. Moreover, the superior performance of this TCN highlights its potential as a promising deep learning model for similar hydrological applications.

Guidance on the construction and selection of relatively simple to complex data-driven models for multi-task streamflow forecasting

Guidance on the construction and selection of relatively simple to complex data-driven models for multi-task streamflow forecasting

Scalable and Interpretable Forecasting of Hydrological Time Series Based on Variational Gaussian Processes

Accurate streamflow forecasting is crucial for effectively managing water resources, particularly in countries like Colombia, where hydroelectric power generation significantly contributes to the national energy grid. Although highly interpretable, traditional deterministic, physically-driven models often suffer from complexity and require extensive parameterization. Data-driven models like Linear Autoregressive (LAR) and Long Short-Term Memory (LSTM) networks offer simplicity and performance but cannot quantify uncertainty. This work introduces Sparse Variational Gaussian Processes (SVGPs) for forecasting streamflow contributions. The proposed SVGP model reduces computational complexity compared to traditional Gaussian Processes, making it highly scalable for large datasets. The methodology employs optimal hyperparameters and shared inducing points to capture short-term and long-term relationships among reservoirs. Training, validation, and analysis of the proposed approach consider the streamflow dataset from 23 geographically dispersed reservoirs recorded during twelve years in Colombia. Performance assessment reveals that the proposal outperforms baseline Linear Autoregressive (LAR) and Long Short-Term Memory (LSTM) models in three key aspects: adaptability to changing dynamics, provision of informative confidence intervals through Bayesian inference, and enhanced forecasting accuracy. Therefore, the SVGP-based forecasting methodology offers a scalable and interpretable solution for multi-output streamflow forecasting, thereby contributing to more effective water resource management and hydroelectric planning.

A novel insight on input variable and time lag selection in daily streamflow forecasting using deep learning models

This study investigates the feasibility of using hybrid models namely Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN)-Gated Recurrent Unit (GRU), for short-to-medium range streamflow forecasting in the Mahanadi River basin in India. The performance of these hybrid models is compared with that of standalone models. It investigates the impact of selected parameters and associated time lags on the model performance and offers valuable insights into the use of hybrid models for runoff simulation. The hybrid CNN-LSTM model proves to be robust in capturing the overall time series and the typical high peak flows in both the correlation-based and constant lag cases. Also, the upstream discharges play a significant role in improving the streamflow forecasting. Furthermore, the consideration of all input variables with a constant time lag equal to the basin lag time may yield better flood forecasts, even in cases where computational resources are limited.

Probabilistic multi-step ahead streamflow forecast based on deep learning

Abstract The use of deep learning methods for fluvial flood forecasting is rapidly gaining traction, offering a promising solution to the challenges associated with accurate yet time-consuming numerical models. This paper presents two physics-inspired deep learning approaches specifically designed for fluvial flood forecasting, each embracing different learning principles: centralized and federated learning. The centralized model utilizes an Encoder-Decoder technique to handle input data of varying types and scales, while the federated model is based on a node-link graph with a seq2seq internal model. Both models are enhanced with a probabilistic forecasting head to account for the inherent uncertainty in streamflow forecasts. The objective of these approaches is to address the limitations of traditional numerical models while leveraging the potential of deep learning to improve the speed, accuracy, and scalability of flood forecasting. To validate their effectiveness, the models were tested across different use cases. The findings from the federated learning approach emphasize the importance of catchment clustering before model utilization and demonstrate the models’ ability to generalize effectively in catchments with similar properties. On the other hand, the results of the centralized method highlight the model’s reliance on the test set falling within the data range of the training set (Average NSE and KGE for the sixth hour ahead of 0.88 and 0.78, respectively). To address this limitation, the paper suggests the development of a method for the future, such as leveraging a numerical model or using Generative Adversarial Networks, to generate highly extreme events, particularly in the context of a changing climate. The models are implemented in a flexible operational framework based on open standards, ensuring their adaptability and usability in various settings.

Enhanced daily streamflow forecasting in Northeastern Algeria: integrating hybrid machine learning with advanced wavelet transformation techniques

Enhanced daily streamflow forecasting in Northeastern Algeria: integrating hybrid machine learning with advanced wavelet transformation techniques

Design and implementation of a BigQuery dataset and application programmer interface (API) for the U.S. National Water Model

We introduce an open-source web-based Application Programming Interface (API) developed within a representational state transfer (REST) architecture framework that provides access to the operational streamflow forecasts from the U.S. National Water Model (NWM). We built this API within the Google Cloud infrastructure, taking advantage of Google's API Gateway, BigQuery, and the Google Cloud Run architecture. We ran a data transformation from netCDF to tabular formation then ingested data into BigQuery using Apache Beam parallel processing technology. The API activates functions deployed to Cloud Run that executes SQL queries within the BigQuery data warehouse. The API gives users granular control to specify queries based on forecast type, reference datetime, stream segment specifics, and forecast ensemble members. This API greatly simplifies access and use of current and historic NWM forecasts, an otherwise arduous task due to the large volume of data and unwieldy storage methods. Retrieved forecast data can be used for many critical water management applications providing actionable intelligence for, e.g., flood management, safety, recreation, agriculture, and related uses.