Univariate streamflow forecasting using commonly used data-driven models: literature review and case study

Eight data-driven models and five data pre-processing methods were summarized; the multiple linear regression (MLR), artificial neural network (ANN) and wavelet decomposition (WD) models were then used in short-term streamflow forecasting at four stations in the East River basin, China. The wavelet–artificial neural network (W-ANN) method was used to predict 1-month-ahead monthly streamflow at Longchuan station (LS). The results indicate better performance of MLR and wavelet–multiple linear regression (W-MLR) in analysing the stationary trained dataset. Four models showed similar performance in 1-day-ahead streamflow forecasting, while W-MLR and W-ANN performed better in 5-day-ahead forecasting. Three reservoirs were shown to have more influence on downstream than upstream streamflow and models had the worst performance at Boluo station. Furthermore, the W-ANN model performed well for 1-month-ahead streamflow forecasting at LS with consideration of a deterministic component.     

Improving flow forecasting by error correction modelling in altered catchment conditions

Despite human is an increasingly significant component of the hydrologic cycle in many river basins, most hydrologic models are still developed to accurately reproduce the natural processes and ignore the effect of human activities on the watershed response. This results in non‐stationary model forecast errors and poor predicting performance every time these models are used in non‐pristine watersheds. In the last decade, the representation of human activities in hydrological models has been extensively studied. However, mathematical models integrating the human and the natural dimension are not very common in hydrological applications and nearly unknown in the day‐to‐day practice. In this paper, we propose a new simple data‐driven flow forecast correction method that can be used to simultaneously tackle forecast errors from structural, parameter and input uncertainty, and errors that arise from neglecting human‐induced alterations in conceptual rainfall–runoff models. The correction system is composed of two layers: (i) a classification system that, based on the current flow condition, detects whether the source of error is natural or human induced and (ii) a set of error correction models that are alternatively activated, each tailored to the specific source of errors. As a case study, we consider the highly anthropized Aniene river basin in Italy, where a flow forecasting system is being established to support the operation of a hydropower dam. Results show that, even by using very basic methods, namely if‐then classification rules and linear correction models, the proposed methodology considerably improves the forecasting capability of the original hydrological model. Copyright © 2013 John Wiley & Sons, Ltd.     

Quantification of parametric uncertainty of ANN models with GLUE method for different streamflow dynamics

This study describes the parametric uncertainty of artificial neural networks (ANNs) by employing the generalized likelihood uncertainty estimation (GLUE) method. The ANNs are used to forecast daily streamflow for three sub-basins of the Rhine Basin (East Alpine, Main, and Mosel) having different hydrological and climatological characteristics. We have obtained prior parameter distributions from 5000 ANNs in the training period to capture the parametric uncertainty and subsequently 125,000 correlated parameter sets were generated. These parameter sets were used to quantify the uncertainty in the forecasted streamflow in the testing period using three uncertainty measures: percentage of coverage, average relative length, and average asymmetry degree. The results indicated that the highest uncertainty was obtained for the Mosel sub-basin and the lowest for the East Alpine sub-basin mainly due to hydro-climatic differences between these basins. The prediction results and uncertainty estimates of the proposed methodology were compared to the direct ensemble and bootstrap methods. The GLUE method successfully captured the observed discharges with the generated prediction intervals, especially the peak flows. It was also illustrated that uncertainty bands are sensitive to the selection of the threshold value for the Nash–Sutcliffe efficiency measure used in the GLUE method by employing the Wilcoxon–Mann–Whitney test.     

Calibration of hydrological models in ungauged basins based on satellite radar altimetry observations of river water level

This paper proposes a new orientation to address the problem of hydrological model calibration in ungauged basin. Satellite radar altimetric observations of river water level at basin outlet are used to calibrate the model, as a surrogate of streamflow data. To shift the calibration objective, the hydrological model is coupled with a hydraulic model describing the relation between streamflow and water stage. The methodology is illustrated by a case study in the Upper Mississippi Basin using TOPEX/Poseidon (T/P) satellite data. The generalized likelihood uncertainty estimation (GLUE) is employed for model calibration and uncertainty analysis. We found that even without any streamflow information for regulating model behavior, the calibrated hydrological model can make fairly reasonable streamflow estimation. In order to illustrate the degree of additional uncertainty associated with shifting calibration objective and identifying its sources, the posterior distributions of hydrological parameters derived from calibration based on T/P data, streamflow data and T/P data with fixed hydraulic parameters are compared. The results show that the main source is the model parameter uncertainty. And the contribution of remote sensing data uncertainty is minor. Furthermore, the influence of removing high error satellite observations on streamflow estimation is also examined. Under the precondition of sufficient temporal coverage of calibration data, such data screening can eliminate some unrealistic parameter sets from the behavioral group. The study contributes to improve streamflow estimation in ungauged basin and evaluate the value of remote sensing in hydrological modeling. Copyright © 2011 John Wiley & Sons, Ltd.     

Forecasting streamflow by combination of a genetic input selection algorithm and wavelet transforms using ANFIS models

ABSTRACT

In this study, a data-driven streamflow forecasting model is developed, in which appropriate model inputs are selected using a binary genetic algorithm (GA). The process involves using a combination of a GA input selection method and two adaptive neuro-fuzzy inference systems (ANFIS): subtractive (Sub)-ANFIS and fuzzy C-means (FCM)-ANFIS. Moreover, the application of wavelet transforms coupled with these models is tested. Long-term data for the Lighvan and Ajichai basins in Iran are used to develop the models. The results indicate considerable improvements when GA selection and wavelet methods are used in both models. For example, the Nash-Sutcliffe efficiency (NSE) coefficient for Lighvan using FCM-ANFIS is 0.74. However, when GA selection is applied, the NSE is improved to 0.85. Moreover, when the wavelet method is added, the performance of new hybrid models shows noticeable enhancements. The NSE value of wavelet-FCM-ANFIS is improved to 0.97 for Lighvan basin.

Editor D. Koutsoyiannis Associate editor E. Toth     

A review of continental scale hydrological models and their suitability for drought forecasting in (sub-Saharan) Africa

The aim of this review is to provide a basis for selecting a suitable hydrological model, or combination of models, for hydrological drought forecasting in Africa at different temporal and spatial scales; for example short and medium range (1–10 days or monthly) forecasts at medium to large river basin scales or seasonal forecasts at the Pan-African scale. Several global hydrological models are currently available with different levels of complexity and data requirements. However, most of these models are likely to fail to properly represent the water balance components that are particularly relevant in arid and semi-arid basins in sub-Saharan Africa. This review critically looks at weaknesses and strengths in the representation of different hydrological processes and fluxes of each model. The major criteria used for assessing the suitability of the models are (1) the representation of the processes that are most relevant for simulating drought conditions, such as interception, evaporation, surface water-groundwater interactions in wetland areas and flood plains and soil moisture dynamics; (2) the capability of the model to be downscaled from a continental scale to a large river basin scale model; and (3) the applicability of the model to be used operationally for drought early warning, given the data availability of the region. This review provides a framework for selecting models for hydrological drought forecasting, conditional on spatial scale, data availability and end-user forecast requirements. Among 16 well known hydrological and land surface models selected for this review, PCR-GLOBWB, GWAVA, HTESSEL, LISFLOOD and SWAT show higher potential and suitability for hydrological drought forecasting in Africa based on the criteria used in this evaluation.     

Improving streamflow estimation in ungauged basins using a multi-modelling approach

ABSTRACT

In this study, a multi-modelling approach is proposed for improved continuous daily streamflow estimation in ungauged basins using regionalization—the process of transferring hydrological data from gauged to ungauged watersheds. Four regionalization models, two data-driven and two hydrological, were used for continuous daily streamflow estimation. Comparison of the individual models reveals that each of the four models performed well on a limited number of ungauged basins while none of them performed well for the entire 90 selected watersheds. The results obtained from the four models are evaluated and reported in a deterministic way by a model combination approach along with its uncertainty range consisting of 16 ensemble members. It is shown that a combined model of the four individual models performed well on all 90 watersheds and the ensemble range can account for the uncertainty of models. The combined model was more efficient and appeared more robust compared to the individual models. Furthermore, continuous ranked probability scores (CRPS) calculated for the ensemble model outputs indicate better performance compared to individual models and competitive with the combined model.

EDITOR A. Castellarin ASSOCIATE EDITOR G. Di Baldassarre     

Structural uncertainty assessment in a discharge simulation model

Abstract

A major goal in hydrological modelling is to identify and quantify different sources of uncertainty in the modelling process. This paper analyses the structural uncertainty in a streamflow modelling system by investigating a set of models with increasing model structure complexity. The models are applied to two basins: Kielstau in Germany and XitaoXi in China. The results show that the model structure is an important factor affecting model performance. For the Kielstau basin, influences from drainage and wetland are critical for the local runoff generation, while for the XitaoXi basin accurate distributions of precipitation and evapotranspiration are two of the determining factors for the success of the river flow simulations. The derived model uncertainty bounds exhibit appropriate coverage of observations. Both case studies indicate that simulation uncertainty for the low-flow period contributes more to the overall uncertainty than that for the peak-flow period, although the main hydrological features in these two basins differ greatly.

Citation Zhang, X. Y., Hörmann, G., Gao, J. F. & Fohrer, N. (2011) Structural uncertainty assessment in a discharge simulation model. Hydrol. Sci. J. 56(5), 854–869.     

Comparison of data-driven modelling techniques for river flow forecasting

Abstract

Accurate forecasting of streamflow is essential for the efficient operation of water resources systems. The streamflow process is complex and highly nonlinear. Therefore, researchers try to devise alterative techniques to forecast streamflow with relative ease and reasonable accuracy, although traditional deterministic and conceptual models are available. The present work uses three data-driven techniques, namely artificial neural networks (ANN), genetic programming (GP) and model trees (MT) to forecast river flow one day in advance at two stations in the Narmada catchment of India, and the results are compared. All the models performed reasonably well as far as accuracy of prediction is concerned. It was found that the ANN and MT techniques performed almost equally well, but GP performed better than both these techniques, although only marginally in terms of prediction accuracy in normal and extreme events.

Citation Londhe, S. & Charhate, S. (2010) Comparison of data-driven modelling techniques for river flow forecasting. Hydrol. Sci. J. 55(7), 1163–1174.     

Use of daily precipitation uncertainties in streamflow simulation and forecast

Among other sources of uncertainties in hydrologic modeling, input uncertainty due to a sparse station network was tested. The authors tested impact of uncertainty in daily precipitation on streamflow forecasts. In order to test the impact, a distributed hydrologic model (PRMS, Precipitation Runoff Modeling System) was used in two hydrologically different basins (Animas basin at Durango, Colorado and Alapaha basin at Statenville, Georgia) to generate ensemble streamflows. The uncertainty in model inputs was characterized using ensembles of daily precipitation, which were designed to preserve spatial and temporal correlations in the precipitation observations. Generated ensemble flows in the two test basins clearly showed fundamental differences in the impact of input uncertainty. The flow ensemble showed wider range in Alapaha basin than the Animas basin. The wider range of streamflow ensembles in Alapaha basin was caused by both greater spatial variance in precipitation and shorter time lags between rainfall and runoff in this rainfall dominated basin. This ensemble streamflow generation framework was also applied to demonstrate example forecasts that could improve traditional ESP (Ensemble Streamflow Prediction) method.     

Assessment of the Long-Term Hydrological Forecast Skill Evolution across Lead-Times within the Ensemble Streamflow Prediction Framework

Long-term or seasonal forecasting is crucial for the management of large water systems. Advances in catchment hydrology, such as mathematical models of catchment processes, are proven to be capable of creating reliable streamflow forecasting systems. In this study, the limits of predictability of streamflow in a snowmelt-dominated river basin are examined and a new illustration of the forecast efficiency across different issue dates and lead times—the so-called “forecastability map”—is demonstrated.     

Influence of spatial distribution of sensors and observation accuracy on the assimilation of distributed streamflow data in hydrological modelling

The aim of this study is to assess the influence of sensor locations and varying observation accuracy on the assimilation of distributed streamflow observations, also taking into account different structures of semi-distributed hydrological models. An ensemble Kalman filter is used to update a semi-distributed hydrological model as a response to measured streamflow. Various scenarios of sensor locations and observation accuracy are introduced. The methodology is tested on the Brue basin during five flood events. The results of this work demonstrate that the assimilation of streamflow observations at interior points of the basin can improve the hydrological models according to the particular location of the sensors and hydrological model structure. It is also found that appropriate definition of the observation accuracy can affect model performance and consequent flood forecasting. These findings can be used as criteria to develop methods for streamflow monitoring network design.     

Assessment of climate change impacts on the hydrology of the Peruvian Amazon–Andes basin

In this article, we propose an investigation of the modifications of the hydrological response of two Peruvian Amazonas–Andes basins in relationship with the modifications of the precipitation and evapotranspiration rates inferred by the IPCC. These two basins integrate around 10% of the total area of the Amazonian basin. These estimations are based on the application of two monthly hydrological models, GR2M and MWB3, and the climatic projections come from BCM2, CSMK3 and MIHR models for A1B and B1 emission scenarios (SCE A1B and SCE B1). Projections are approximated by two simple scenarios (anomalies and horizon) and annual rainfall rates, evapotranspiration rates and discharge were estimated for the 2020s (2008–2040), 2050s (2041–2070) and 2080s (2071–2099). Annual discharge shows increasing trend over Requena basin (Ucayali river), Puerto Inca basin (Pachitea river), Tambo basin (Tambo river) and Mejorada basin (Mantaro river) while discharge shows decreasing trend over the Chazuta basin (Huallaga river), the Maldonadillo basin (Urubamba river) and the Pisac basin (Vilcanota river). Monthly discharge at the outlet of Puerto Inca, Tambo and Mejorada basins shows increasing trends for all seasons. Trends to decrease are estimated in autumn discharge over the Requena basin and spring discharge over Pisac basin as well as summer and autumn discharges over both the Chazuta and the Maldonadillo basins. Copyright © 2011 John Wiley & Sons, Ltd.     

Flood frequency under the influence of trends in the Pearl River basin,China: changing patterns,causes and implications

Using a nonstationary flood frequency model, this study investigates the impact of trends on the estimation of flood frequencies and flood magnification factors. Analysis of annual peak streamflow data from 28 hydrological stations across the Pearl River basin, China, shows that: (1) northeast parts of the West and the North River basins are dominated by increasing annual peak streamflow, whereas decreasing trends of annual peak streamflow are prevailing in other regions of the Pearl River basin; (2) trends significantly impact the estimation of flood frequencies. The changing frequency of the same flood magnitude is related to the changing magnitude or significance/insignificance of trends, larger increasing frequency can be detected for stations with significant increasing trends of annual peak streamflow and vice versa, and smaller increasing magnitude for stations with not significant increasing annual peak streamflow, pointing to the critical impact of trends on estimation of flood frequencies; (3) larger‐than‐1 flood magnification factors are observed mainly in the northeast parts of the West River basin and in the North River basin, implying magnifying flood processes in these regions and a higher flood risk in comparison with design flood‐control standards; and (4) changes in hydrological extremes result from the integrated influence of human activities and climate change. Generally, magnifying flood regimes in the northeast Pearl River basin and in the North River basin are mainly the result of intensifying precipitation regime; smaller‐than‐1 flood magnification factors along the mainstream of the West River basin and also in the East River basin are the result of hydrological regulations of water reservoirs. Copyright © 2014 John Wiley & Sons, Ltd.     

Predictions in ungauged basins: an approach for regionalization of hydrological models considering the probability distribution of model parameters

Regionalization of model parameters by developing appropriate functional relationship between the parameters and basin characteristics is one of the potential approaches to employ hydrological models in ungauged basins. While this is a widely accepted procedure, the uniqueness of the watersheds and the equifinality of parameters bring lot of uncertainty in the simulations in ungauged basins. This study proposes a method of regionalization based on the probability distribution function of model parameters, which accounts the variability in the catchment characteristics. It is envisaged that the probability distribution function represents the characteristics of the model parameter, and when regionalized the earlier concerns can be addressed appropriately. The method employs probability distribution of parameters, derived from gauged basins, to regionalize by regressing them against the catchment attributes. These regional functions are used to develop the parameter characteristics in ungauged basins based on the catchment attributes. The proposed method is illustrated using soil water assessment tool model for an ungauged basin prediction. For this numerical exercise, eight different watersheds spanning across different climatic settings in the USA are considered. While all the basins considered in this study were gauged, one of them was assumed to be ungauged (pseudo-ungauged) in order to evaluate the effectiveness of the proposed methodology in ungauged basin simulation. The process was repeated by considering representative basins from different climatic and landuse scenarios as pseudo-ungauged. The results of the study indicated that the ensemble simulations in the ungauged basins were closely matching with the observed streamflow. The simulation efficiency varied between 57 and 61 % in ungauged basins. The regional function was able to generate the parameter characteristics that were closely matching with the original probability distribution derived from observed streamflow data.     

High-resolution simulation of the spatial pattern of water use in continental China

ABSTRACT

High-resolution data on the spatial pattern of water use are a prerequisite for appropriate and sustainable water management. Based on one well-validated hydrological model, the Distributed Time Variant Gains Model (DTVGM), this paper obtains reliable high-resolution spatial patterns of irrigation, industrial and domestic water use in continental China. During the validation periods, ranges of correlation coefficient (R) and Nash-Sutcliffe efficiency (NSE) coefficient are 0.67–0.96 and 0.51–0.84, respectively, between the observed and simulated streamflow of six hydrological stations, indicating model applicability to simulate the distribution of water use. The simulated water use quantities have relative errors (RE) less than 5% compared with the observed. In addition, the changes in streamflow discharge were also correctly simulated by our model, such as the Zhangjiafen station in the Hai River basin with a dramatic decrease in streamflow, and the Makou station in the Pearl River basin with no significant changes. These changes are combined results of basin available water resources and water use. The obtained high-resolution spatial pattern of water use could decrease uncertainty of hydrological simulation and guide water management efficiently.

Editor M.C. Acreman; Associate editor X. Fang     

Comparison of daily and sub-daily SWAT models for daily streamflow simulation in the Upper Huai River Basin of China

Despite the significant role of precipitation in the hydrological cycle, few studies have been conducted to evaluate the impacts of the temporal resolution of rainfall inputs on the performance of SWAT (soil and water assessment tool) models in large-sized river basins. In this study, both daily and hourly rainfall observations at 28 rainfall stations were used as inputs to SWAT for daily streamflow simulation in the Upper Huai River Basin. Study results have demonstrated that the SWAT model with hourly rainfall inputs performed better than the model with daily rainfall inputs in daily streamflow simulation, primarily due to its better capability of simulating peak flows during the flood season. The sub-daily SWAT model estimated that 58 % of streamflow was contributed by baseflow compared to 34 % estimated by the daily model. Using the future daily and 3-h precipitation projections under the RCP (Representative Concentration Pathways) 4.5 scenario as inputs, the sub-daily SWAT model predicted a larger amount of monthly maximum daily flow during the wet years than the daily model. The differences between the daily and sub-daily SWAT model simulation results indicated that temporal rainfall resolution could have much impact on the simulation of hydrological process, streamflow, and consequently pollutant transport by SWAT models. There is an imperative need for more studies to examine the effects of temporal rainfall resolution on the simulation of hydrological and water pollutant transport processes by SWAT in river basins of different environmental conditions.     

Assessing the relative importance of parameter and forcing uncertainty and their interactions in conceptual hydrological model simulations

Predictions of river flow dynamics provide vital information for many aspects of water management including water resource planning, climate adaptation, and flood and drought assessments. Many of the subjective choices that modellers make including model and criteria selection can have a significant impact on the magnitude and distribution of the output uncertainty. Hydrological modellers are tasked with understanding and minimising the uncertainty surrounding streamflow predictions before communicating the overall uncertainty to decision makers. Parameter uncertainty in conceptual rainfall-runoff models has been widely investigated, and model structural uncertainty and forcing data have been receiving increasing attention. This study aimed to assess uncertainties in streamflow predictions due to forcing data and the identification of behavioural parameter sets in 31 Irish catchments. By combining stochastic rainfall ensembles and multiple parameter sets for three conceptual rainfall-runoff models, an analysis of variance model was used to decompose the total uncertainty in streamflow simulations into contributions from (i) forcing data, (ii) identification of model parameters and (iii) interactions between the two. The analysis illustrates that, for our subjective choices, hydrological model selection had a greater contribution to overall uncertainty, while performance criteria selection influenced the relative intra-annual uncertainties in streamflow predictions. Uncertainties in streamflow predictions due to the method of determining parameters were relatively lower for wetter catchments, and more evenly distributed throughout the year when the Nash-Sutcliffe Efficiency of logarithmic values of flow (lnNSE) was the evaluation criterion.     

Comparison of local and global approximators in multivariate chaotic forecasting of daily streamflow

ABSTRACT

Although it is conceptually assumed that global models are relatively ineffective in modelling the highly unstable structure of chaotic hydrologic dynamics, there is not a detailed study of comparing the performances of local and global models in a hydrological context, especially with new emerging machine learning models. In this study, the performance of a local model (k-nearest neighbour, k-nn) and, as global models, several recent machine learning models – artificial neural network (ANN), least square-support vector regression (LS-SVR), random forest (RF), M5 model tree (M5), multivariate adaptive regression splines (MARS) – was analysed in multivariate chaotic forecasting of streamflow. The models were developed for Australia’s largest river, the River Murray. The results indicate that the k-nn model was more successful than the global models in capturing the streamflow dynamics. Furthermore, coupled with the multivariate phase-space, it was shown that the global models can be successfully used for obtaining reliable uncertainty estimates for streamflow.     

Application of tank,NAM, ARMA and neural network models to flood forecasting

Two lumped conceptual hydrological models, namely tank and NAM and a neural network model are applied to flood forecasting in two river basins in Thailand, the Wichianburi on the Pasak River and the Tha Wang Pha on the Nan River using the flood forecasting procedure developed in this study. The tank and NAM models were calibrated and verified and found to give similar results. The results were found to improve significantly by coupling stochastic and deterministic models (tank and NAM) for updating forecast output. The neural network (NN) model was compared with the tank and NAM models. The NN model does not require knowledge of catchment characteristics and internal hydrological processes. The training process or calibration is relatively simple and less time consuming compared with the extensive calibration effort required by the tank and NAM models. The NN model gives good forecasts based on available rainfall, evaporation and runoff data. The black‐box nature of the NN model and the need for selecting parameters based on trial and error or rule‐of‐thumb, however, characterizes its inherent weakness. The performance of the three models was evaluated statistically. Copyright © 2000 John Wiley & Sons, Ltd.