Short-term Streamflow Forecasting Research Articles

Comparative evaluation of daily streamflow prediction by ANN and SWAT models in two karst watersheds in central south Texas

Abstract This work compares the accuracy of streamflow estimated by a data-driven artificial neural network (ANN) and the physically based soil and water assessment tool (SWAT). The models were applied in two small watersheds, one highly urbanized and the other primarily covered with evergreen forest and shrubs, in the San Antonio region of central south Texas, where karst geologic features are prevalent. Both models predicted daily streamflow in the urbanized watershed very well, with the ANN and SWAT having the Nash–Sutcliffe coefficient of efficiency (NSE) values of 0.76 and 0.72 in the validation period, respectively. However, both models predicted streamflow poorly in the nonurban watershed. The NSE values of the ANNs significantly improved when a time series autoregressive model structure using historical streamflow data was implemented in the nonurban watershed. The SWAT model only achieved trivial performance improvement after using the SWAT-CUP SUFI-2 calibration procedure. This result suggests that an ANN model may be more suitable for short-term streamflow forecasting in watersheds heavily affected by karst features where the complex processes of rapid groundwater recharge and discharge strongly influence surface water flow.

Comparative evaluation of LSTM, CNN, and ConvLSTM for hourly short-term streamflow forecasting using deep learning approaches

This study investigates the effectiveness of three deep learning methods, Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Convolutional Long Short-Term Memory (ConvLSTM), in short-term streamflow forecasting in the Kelantan and Muda River basins in Malaysia. The ability to forecast streamflow in the short-term is crucial for water resource management to prevent disasters such as floods. Hourly streamflow and rainfall data from the two river basins were used to evaluate the accuracy of the methods at one-hour, three-hour, and six-hour prediction intervals. The results showed that all three deep learning methods performed with high accuracy in predicting streamflow. The study also found that LSTM performed better in small basin with well spatial distributed rainfall stations, while CNN and ConvLSTM were more effective in moderate to high streamflow, and large river basin. Moreover, early prediction intervals had better accuracy compared to later ones. The findings of this study demonstrate the potential of using deep learning methods for short-term streamflow forecasting in water resource management. The study highlights the importance of accurate short-term streamflow forecasting and emphasizes the potential of using deep learning algorithms to achieve this goal.

Short-term forecasts of streamflow in the UK based on a novel hybrid artificial intelligence algorithm

In recent years, the growing impact of climate change on surface water bodies has made the analysis and forecasting of streamflow rates essential for proper planning and management of water resources. This study proposes a novel ensemble (or hybrid) model, based on the combination of a Deep Learning algorithm, the Nonlinear AutoRegressive network with eXogenous inputs, and two Machine Learning algorithms, Multilayer Perceptron and Random Forest, for the short-term streamflow forecasting, considering precipitation as the only exogenous input and a forecast horizon up to 7 days. A large regional study was performed, considering 18 watercourses throughout the United Kingdom, characterized by different catchment areas and flow regimes. In particular, the predictions obtained with the ensemble Machine Learning-Deep Learning model were compared with the ones achieved with simpler models based on an ensemble of both Machine Learning algorithms and on the only Deep Learning algorithm. The hybrid Machine Learning-Deep Learning model outperformed the simpler models, with values of R2 above 0.9 for several watercourses, with the greatest discrepancies for small basins, where high and non-uniform rainfall throughout the year makes the streamflow rate forecasting a challenging task. Furthermore, the hybrid Machine Learning-Deep Learning model has been shown to be less affected by reductions in performance as the forecasting horizon increases compared to the simpler models, leading to reliable predictions even for 7-day forecasts.

МЕТОДИКА РАСЧЕТА СРЕДНЕСУТОЧНЫХ РАСХОДОВ ВОДЫ НА РЕКАХ РОССИИ С ИСПОЛЬЗОВАНИЕМ МОДЕЛИ ФОРМИРОВАНИЯ СТОКА HBV-96

A method for calculating the daily streamflow was developed. The methodology uses the HBV-96 conceptual hydrological model. An output adjustment method was applied to the model calculations, which made it possible to significantly improve overall accuracy. The software package was created to develop, implement, and verify the methodology. The technique was implemented for 590 river catchments located throughout Russia in various natural conditions. An archive of all necessary hydrometeorological data for the period from 2010 to 2019 was formed. The results of verification based on an independent sample indicated a fairly high accuracy of calculations for the vast majority of river catchments. The results provide a basis for the development of an automated system for generating and issuing short-term streamflow forecasts for Russian rivers based on the proposed methodology.

Improving short-term streamflow forecasting by flow mode clustering

Improving short-term streamflow forecasting by flow mode clustering

Optimal cropping pattern based on short-term streamflow forecasts to improve agricultural economic benefits and crop productivity under uncertainty conditions

ABSTRACT Accelerated urbanization has led to diminished land for agricultural activities. Riverine ecosystems play an important role in allocating fertile lands to support agricultural activities. A substantial component of the uncertainty in agricultural productivity comes from seasonal variations linked to inter-annual climate fluctuations. Therefore, understanding the complicated phenomena of streamflow in a riverine environment is important for agricultural and water resources decision making. The present work focuses on forecasting monthly to seasonal streamflow using persistence flow, historical analogues, and artificial neural network approaches. Based on these forecasts, decisions on cropping patterns were made by developing an optimization framework using the constrained linear programming and inexact multiobjective fuzzy linear programming approaches. The proposed fuzzy programming approach was found to be beneficial in producing fair and stable solutions under uncertainty. The findings reveal that integrating forecasting and optimization knowledge could aid in precisely evaluating ecosystem services and meeting rising food demand.

Short-Term Hydrological Forecast Using Artificial Neural Network Models with Different Combinations and Spatial Representations of Hydrometeorological Inputs

In hydrological modelling, artificial neural network (ANN) models have been popular in the scientific community for at least two decades. The current paper focuses on short-term streamflow forecasting, 1 to 7 days ahead, using an ANN model in two northeastern American watersheds, the Androscoggin and Susquehanna. A virtual modelling environment is implemented, where data used to train and validate the ANN model were generated using a deterministic distributed model over 16 summers (2000–2015). To examine how input variables affect forecast accuracy, we compared streamflow forecasts from the ANN model using four different sets of inputs characterizing the watershed state—surface soil moisture, deep soil moisture, observed streamflow the day before the forecast, and surface soil moisture along with antecedent observed streamflow. We found that the best choice of inputs consists of combining surface soil moisture with observed streamflow for the two watersheds under study. Moreover, to examine how the spatial distribution of input variables affects forecast accuracy, we compared streamflow forecasts from the ANN using surface soil moisture at three spatial distributions—global, fully distributed, and single pixel-based—for the Androscoggin watershed. We show that model performance was similar for both the global and fully distributed representation of soil moisture; however, both models surpass the single pixel-based models. Future work includes evaluating the developed ANN model with real observations, quantified in situ or remotely sensed.

Quantifying streamflow predictive uncertainty for the optimization of short-term cascade hydropower stations operations

Quantifying the impact of the streamflow forecast uncertainty on the short-term (referring to the next day) power generation operation of a cascade hydropower system is more complex than a single reservoir but important, due to the compensation effect between upstream and downstream reservoirs within a cascade. In this paper, we propose a short-term cascade hydropower stations optimal operation model based on conditional value-at-risk (called the CVaR-SCHOM model) to guide the actual power generation using uncertain streamflow forecast. We first analyze the uncertainty of short-term streamflow forecasts and establish the conditional probability function to describe the uncertainty of the forecasted streamflow sequence. We next propose the short-term operation target of the cascade hydropower station at the greatest expected benefit. We then introduce the conditional value-at-risk theory to analyze the risk of power shortage in different operating schemes under a certain level of confidence. Finally, based on the risk attitude of the decision-makers, we balance the benefit and risk of cascade hydropower stations operation and establish the CVaR-SCHOM model. We verify the model using the Jinguan cascade hydropower stations in the Yalung River basin. The main contributions are as follows. 1) By using k-means clustering to classify the forecasted streamflow sequence into different magnitudes and then using the joint probability function of the relative errors, an accurate description of the uncertainty of the short-term streamflow forecast can be achieved. 2) The internal characteristics of each cascade power station and the interrelationships between the cascade hydropower stations are complicated, and the power deviation and forecast error can show positive or negative correlations. 3) Compared with the traditional deterministic optimization model, the CVaR-SCHOM proposed in this paper performs well in real-world applications. The proposed model yields different degrees of improvement of the benefits and risks for various forecast magnitudes. The decision-makers can choose different schemes according to their risk attitudes.

Evaluation of random forests for short-term daily streamflow forecasting in rainfall- and snowmelt-driven watersheds

Abstract. In the past decades, data-driven machine-learning (ML) models have emerged as promising tools for short-term streamflow forecasting. Among other qualities, the popularity of ML models for such applications is due to their relative ease in implementation, less strict distributional assumption, and competitive computational and predictive performance. Despite the encouraging results, most applications of ML for streamflow forecasting have been limited to watersheds in which rainfall is the major source of runoff. In this study, we evaluate the potential of random forests (RFs), a popular ML method, to make streamflow forecasts at 1 d of lead time at 86 watersheds in the Pacific Northwest. These watersheds cover diverse climatic conditions and physiographic settings and exhibit varied contributions of rainfall and snowmelt to their streamflow. Watersheds are classified into three hydrologic regimes based on the timing of center-of-annual flow volume: rainfall-dominated, transient, and snowmelt-dominated. RF performance is benchmarked against naïve and multiple linear regression (MLR) models and evaluated using four criteria: coefficient of determination, root mean squared error, mean absolute error, and Kling–Gupta efficiency (KGE). Model evaluation scores suggest that the RF performs better in snowmelt-driven watersheds compared to rainfall-driven watersheds. The largest improvements in forecasts compared to benchmark models are found among rainfall-driven watersheds. RF performance deteriorates with increases in catchment slope and soil sandiness. We note disagreement between two popular measures of RF variable importance and recommend jointly considering these measures with the physical processes under study. These and other results presented provide new insights for effective application of RF-based streamflow forecasting.

Identification of Combined Hydrological Models and Numerical Weather Predictions for Enhanced Flood Forecasting in a Semiurban Watershed

Flood forecasting in urban and semiurban catchments is often limited by the capability of the combined hydrological models and forecast inputs to predict floods accurately. The objective of this research is to develop an approach (1) to identify the best model forecast from multiple integrations of various hydrological models and numerical weather predictions (NWP), and (2) to find the best forecast combination method for an improved short-range flood forecasting. Seven selected hydrological models were coupled, each with two high-resolution NWP forecasts to provide several alternatives of deterministic hydrological forecasts at a catchment outlet. As such, the different model-input combinations were used to generate 14 hydrological forecasts. Hydrological forecast verification was then carried out over a one-year hindcast period. A comparison between six forecast combination methods, including a benchmark Bayesian model averaging (BMA) method, was also performed for the multiple available short-term streamflow forecasts. Results indicate that the coupling of the Sacramento soil moisture accounting (SACSMA) model with both High-Resolution Deterministic Precipitation System and High-Resolution Rapid Refresh inputs outperformed other model-input integrations. Maximum forecast errors in all model-input integration outputs occurred at forecast lead times of 12–14 h, corresponding to the time of concentration of the catchment. Providing constraints on the estimation of model weights was found to be a significant factor for obtaining an improved combined streamflow forecast. In general, the regression-based forecast combination method of the constrained ordinary least squares (CLS) has emerged as a possible alternative to the widely used BMA method for hydrology application.

A comparative study of models for short-term streamflow forecasting with emphasis on wavelet-based approach

Skilful short-term streamflow forecasting is a challenging task, but useful for addressing a variety of issues associated with water resources planning and management. This study investigates the performance of four popular models for short-term streamflow forecasting, with particular emphasis on the combination of each model with wavelet transform (WT). The autoregressive (AR) model, autoregressive moving average (ARMA) model, artificial neural network (ANN) model, and linear regression (LR) model are used as the base models, and WT-AR, WT-ARMA, WT-ANN, and WT-LR models are used as the composite or hybrid models. These eight models are applied for short-term forecasting of daily streamflow time series in two different river basins in China. Streamflow data from two stations (Yingluoxia and Zhamashike) in the Heihe River basin of North China and two stations (Wuzhou and Longchuan) in the Pearl River basin of South China are considered, and forecasts are made 1-day, 2-day, and 3-day ahead. The accuracy of the models is evaluated using three widely used performance measures: root mean square error, mean absolute error, and correlation coefficient (R). The performances of the eight models are compared, and the influencing role of the wavelet transform in the hybrid models is discussed. The results suggest that the WT-ANN and ANN models are more suitable for the northern basin but perform poorly for the southern basin, while the WT-ARMA model is more suitable for the southern basin. The results also suggest that the wavelet-based hybrid models are better than the single models particularly for longer lead times, offering gradual improvement with increasing lead time. Therefore, the wavelet-based models are especially useful for areas that suffer from frequent natural disasters, fragile ecological environment, and poor self-regulations.

Analysis of the use of discrete wavelet transforms coupled with ANN for short-term streamflow forecasting

The use of wavelet transforms to forecast daily streamflows into the Sobradinho Reservoir (Bahia State, Brazil) seven days ahead by a wavelet-artificial neural network (ANN) hybrid system was analyzed in this paper. This work also determined the appropriate mother-wavelet for this type of forecasting with an ANN, performed 1836 simulations with the wavelet-ANN hybrid systems (tested with 54 mother-wavelets) and compared the results with the predictions made without the application of a wavelet transform (henceforth called a stand-alone ANN). Daily data from January 1931 to December 2010 were used. According to the results, the wavelet-ANN hybrid system performed better than the system using the ANN with the raw data. The approximation A3 from the discrete Meyer mother-wavelet obtained the best results; the root mean square error (RMSE) decreased by approximately 80%, while the R2 and NASH coefficients increased by more than 5% and 10%, respectively, compared with the stand-alone ANN.

Forecasting Turbidity during Streamflow Events for Two Mid-Atlantic U.S. Streams

Short-term streamflow forecasting is a widely used and important aspect of modern water management. In contrast, routine operational forecasting of stream water quality remains relatively limited. Turbidity is a commonly-monitored, key water-quality parameter. It can often be used to estimate other water-quality parameters and can serve as an overall indicator of stream environmental health. In this study, short-term (3-day) turbidity forecasts during streamflow events for two Mid-Atlantic U.S. streams were produced using a combination of forecast discharge, precipitation and air temperature, together with observations leading up to the issue time of the forecast. The turbidity forecast error was found to be relatively constant with lead time and significantly less than the persistence reference error for nearly all lead times. The turbidity forecast uncertainty due to streamflow forecast uncertainty was also evaluated. Potential future improvements for the example turbidity forecasts presented here are discussed. This study demonstrates for the first time that currently-available inputs (i.e., forecast discharge, precipitation and air temperature) can yield useful stream turbidity forecasts.

A retrospective streamflow ensemble forecast for an extreme hydrologic event: a case study of Hurricane Irene and on the Hudson River basin

Abstract. This paper investigates the uncertainties in hourly streamflow ensemble forecasts for an extreme hydrological event using a hydrological model forced with short-range ensemble weather prediction models. A state-of-the art, automated, short-term hydrologic prediction framework was implemented using GIS and a regional scale hydrological model (HEC-HMS). The hydrologic framework was applied to the Hudson River basin ( ∼ 36 000 km2) in the United States using gridded precipitation data from the National Centers for Environmental Prediction (NCEP) North American Regional Reanalysis (NARR) and was validated against streamflow observations from the United States Geologic Survey (USGS). Finally, 21 precipitation ensemble members of the latest Global Ensemble Forecast System (GEFS/R) were forced into HEC-HMS to generate a retrospective streamflow ensemble forecast for an extreme hydrological event, Hurricane Irene. The work shows that ensemble stream discharge forecasts provide improved predictions and useful information about associated uncertainties, thus improving the assessment of risks when compared with deterministic forecasts. The uncertainties in weather inputs may result in false warnings and missed river flooding events, reducing the potential to effectively mitigate flood damage. The findings demonstrate how errors in the ensemble median streamflow forecast and time of peak, as well as the ensemble spread (uncertainty) are reduced 48 h pre-event by utilizing the ensemble framework. The methodology and implications of this work benefit efforts of short-term streamflow forecasts at regional scales, notably regarding the peak timing of an extreme hydrologic event when combined with a flood threshold exceedance diagram. Although the modeling framework was implemented on the Hudson River basin, it is flexible and applicable in other parts of the world where atmospheric reanalysis products and streamflow data are available.

The application of Dynamic Linear Bayesian Models in hydrological forecasting: Varying Coefficient Regression and Discount Weighted Regression

Summary A novel implementation of Dynamic Linear Bayesian Models (DLBM), using either a Varying Coefficient Regression (VCR) or a Discount Weighted Regression (DWR) algorithm was used in the hydrological modeling of annual hydrographs as well as 1-, 2-, and 3-day lead time stream flow forecasting. Using hydrological data (daily discharge, rainfall, and mean, maximum and minimum air temperatures) from the Upper Narew River watershed in Poland, the forecasting performance of DLBM was compared to that of traditional multiple linear regression (MLR) and more recent artificial neural network (ANN) based models. Model performance was ranked DLBM-DWR > DLBM-VCR > MLR > ANN for both annual hydrograph modeling and 1-, 2-, and 3-day lead forecasting, indicating that the DWR and VCR algorithms, operating in a DLBM framework, represent promising new methods for both annual hydrograph modeling and short-term stream flow forecasting.

Joint and respective effects of long- and short-term forecast uncertainties on reservoir operations

Summary Forecast uncertainties are a major problem in forecast-based water resource management. This study aims to improve the combined use of long- and short-term streamflow forecasts in reservoir operations by investigating the joint and respective effects of long- and short-term forecast uncertainties on decision-making processes. Forecast uncertainties in ensemble streamflow forecasts are explained, and the joint effect is characterized by a matrix of operation decisions. A novel statistical partitioning method is developed to diagnose the respective effects of long- and short-term forecast uncertainties, particularly on an experimental operation of the Danjiangkou Reservoir in China. Given the large regulating capacity of this reservoir, long-term forecast uncertainties inevitably surround its release decisions. When a fixed carried-over storage is used to couple long- and short-term forecasts, an ending storage effect is observed. As a stage approaches its end, the release decision becomes increasingly dependent on the carried-over storage, thereby increasing uncertainty in decision-making as a result of long-term forecast uncertainty. This finding highlights the importance of developing reliable long-term forecasts and selecting a proper carried-over storage to aid the decision-making process. This study proposes a sliding carried-over storage strategy to circumvent the ending storage effect. Results show that, using the same long-term forecast, the proposed strategy significantly alleviates the ending storage effect and reduces uncertainties in the decision-making process. The sliding carried-over storage strategy provides an efficient approach to the combined use of long- and short-term forecasts and improves reservoir operations.

Post-processing rainfall forecasts from numerical weather prediction models for short-term streamflow forecasting

Abstract. Sub-daily ensemble rainfall forecasts that are bias free and reliably quantify forecast uncertainty are critical for flood and short-term ensemble streamflow forecasting. Post-processing of rainfall predictions from numerical weather prediction models is typically required to provide rainfall forecasts with these properties. In this paper, a new approach to generate ensemble rainfall forecasts by post-processing raw numerical weather prediction (NWP) rainfall predictions is introduced. The approach uses a simplified version of the Bayesian joint probability modelling approach to produce forecast probability distributions for individual locations and forecast lead times. Ensemble forecasts with appropriate spatial and temporal correlations are then generated by linking samples from the forecast probability distributions using the Schaake shuffle. The new approach is evaluated by applying it to post-process predictions from the ACCESS-R numerical weather prediction model at rain gauge locations in the Ovens catchment in southern Australia. The joint distribution of NWP predicted and observed rainfall is shown to be well described by the assumed log-sinh transformed bivariate normal distribution. Ensemble forecasts produced using the approach are shown to be more skilful than the raw NWP predictions both for individual forecast lead times and for cumulative totals throughout all forecast lead times. Skill increases result from the correction of not only the mean bias, but also biases conditional on the magnitude of the NWP rainfall prediction. The post-processed forecast ensembles are demonstrated to successfully discriminate between events and non-events for both small and large rainfall occurrences, and reliably quantify the forecast uncertainty. Future work will assess the efficacy of the post-processing method for a wider range of climatic conditions and also investigate the benefits of using post-processed rainfall forecasts for flood and short-term streamflow forecasting.

Evaluation of numerical weather prediction model precipitation forecasts for short-term streamflow forecasting purpose

Abstract. The quality of precipitation forecasts from four Numerical Weather Prediction (NWP) models is evaluated over the Ovens catchment in Southeast Australia. Precipitation forecasts are compared with observed precipitation at point and catchment scales and at different temporal resolutions. The four models evaluated are the Australian Community Climate Earth-System Simulator (ACCESS) including ACCESS-G with a 80 km resolution, ACCESS-R 37.5 km, ACCESS-A 12 km, and ACCESS-VT 5 km. The skill of the NWP precipitation forecasts varies considerably between rain gauging stations. In general, high spatial resolution (ACCESS-A and ACCESS-VT) and regional (ACCESS-R) NWP models overestimate precipitation in dry, low elevation areas and underestimate in wet, high elevation areas. The global model (ACCESS-G) consistently underestimates the precipitation at all stations and the bias increases with station elevation. The skill varies with forecast lead time and, in general, it decreases with the increasing lead time. When evaluated at finer spatial and temporal resolution (e.g. 5 km, hourly), the precipitation forecasts appear to have very little skill. There is moderate skill at short lead times when the forecasts are averaged up to daily and/or catchment scale. The precipitation forecasts fail to produce a diurnal cycle shown in observed precipitation. Significant sampling uncertainty in the skill scores suggests that more data are required to get a reliable evaluation of the forecasts. The non-smooth decay of skill with forecast lead time can be attributed to diurnal cycle in the observation and sampling uncertainty. Future work is planned to assess the benefits of using the NWP rainfall forecasts for short-term streamflow forecasting. Our findings here suggest that it is necessary to remove the systematic biases in rainfall forecasts, particularly those from low resolution models, before the rainfall forecasts can be used for streamflow forecasting.

Application of artificial neural networks and adaptive neuro-fuzzy inference system models to short-term streamflow forecasting

The present article aims to forecast streamflow by using artificial neural networks (ANNs), adaptive neuro-fuzzy inference system (ANFIS), and autoregressive moving average (ARMA). For this purpose, the daily streamflow time series of two hydrometry stations of Hajighoshan and Tamar on Gorgan River are used for two periods of 1983–2007 and 1974–2007, respectively. Root mean square error (RMSE) and correlation coefficient (R) statistics are employed to evaluate the performance of the ANNs, ANFIS, and ARMA models for forecasting streamflow (1 day ahead). Comparison of the results reveals that the ANFIS model outperforms the ARMA model. Based on the results of validation stage, for the forecasting 1 day ahead streamflow, ANN with RMSE = 0.028 m3/s and R = 0.59 for the Hajighoshan station and RMSE = 0.013 m3/s and R = 0.44 for the Tamar station were found to be superior to the ANFIS with RMSE = 1.98 m3/s and R = 0.42 for the Hajighoshan station and RMSE = 2.18 m3/s and R = 0.22 for the Tamar station. In addition, for 2 day and 3 day ahead streamflow forecasts, the ANN models show superiority in the accuracy of forecasting streamflow compared with the ANFIS models.

Wavelet regression model for short-term streamflow forecasting

Summary Wavelet regression (WR) technique is proposed for short-term streamflow forecasting in this study. The WR model is improved combining two methods, discrete wavelet transform and linear regression model. The proposed model is applied to the daily streamflow data of two stations, Karabuk and Derecikviran, on the Filyos River in the Western Black Sea region of Turkey. Root mean square error (RMSE), mean absolute error (MAE) and correlation coefficient (R) statistics are used for evaluating the accuracy of the WR models. In the first part of the study, the accuracy of the WR models is compared with the artificial neural network (ANN) and autoregressive moving average (ARMA) models in 1-day ahead streamflow forecasting. Comparison results reveal that the WR model performs better than the ANN and ARMA models. The ARMA model is also found to be slightly better than the ANN. For the Karabuk and Derecikviran stations, it was found that WR models with RMSE = 8.48 m3/s, MAE = 2.46 m3/s, R = 0.978 and RMSE = 33.3 m3/s, MAE = 10.2 m3/s, R = 0.976 in the validation stage are superior in forecasting 1-day ahead streamflows than the most accurate ARMA models with RMSE = 13.5 m3/s, MAE = 3.44 m3/s, R = 0.942 and RMSE = 46.5 m3/s, MAE = 13.2 m3/s, R = 0.953, respectively. In the second part of study, WR and ANN models are compared in 2- and 3-day ahead streamflow forecasting. Based on the comparison results, WR models are found to be more accurate than the ANN models.