Accuracy Of Streamflow Prediction Research Articles

Developing an ANN Based Streamflow Forecast Model Utilizing Data-Mining Techniques to Improve Reservoir Streamflow Prediction Accuracy: A Case Study

This study illustrates the benefits of data pre-processing through supervised data-mining techniques and utilizing those processed data in an artificial neural networks (ANNs) for streamflow prediction. Two major categories of physical parameters such as snowpack data and time-dependent trend indices were utilized as predictors of streamflow values. Correlation analysis of different models indicate that, for the period of January to June, using fewer predictors led to simpler modeling with equivalent accuracy on daily prediction models. This did not hold in all periods. For monthly prediction models, accuracy was improved compared to earlier works done to predict monthly streamflow for the same case of Elephant Butte Reservoir (EB), NM. Overall, superior prediction performance was achieved by utilizing data-mining techniques for pre-processing historical data, extracting the most effective predictors, correlation analysis, extracting and utilizing combined climate variability indices, physical indices, and employing several developed ANNs for different prediction periods of the year.

GIS-Based Rainfall-Runoff Neuro Model for Streamflow Prediction

TOPMODEL is a semi-distributed rainfall runoff model that has been widely used in numerous water resources’ applications in the last few decades. However, literature has identified the weakness in the TOPMODEL performances in streamflow prediction. In this paper, a multilayer perceptron neural network (MLP-NN) has been adopted to improve the accuracy of streamflow prediction in a flash flood in Pinang catchment area. Two daily hydro-meteorological datasets of year 2007-2008 and 2009-2010 were used for calibration and validation periods, respectively. The new method presented in this study uses the TOPMODEL input-output datasets during the calibration period to train the MLP-NN to predict the output. Then, the trained MLP-NN model structure is used to predict the streamflow based on validation period datasets. The three efficiencies considered to evaluate the model performances are the Nash-Sutcliffe model (NS), the Relative Volume Error (RVE) and the Correlation Coefficient (CoC). The results indicated an improvement from 0.749, -19.2 and 0.893 of NS, RVE and CoC of the calibration period to 0.978, 0.364 and 0.989, respectively. Moreover, for the validation periods, the performance has been improved from 0.774, -19.84 and 0.933 of NS, RVE and CoC to 0.975, -0.029 and 0.789, respectively. The ability of MLP-NN to improve TOPMODEL daily streamflow prediction has been demonstrated in this study.

Rainfall-runoff modeling at Jinsha River basin by integrated neural network with discrete wavelet transform

Artificial neural network (ANN) models combined with time series decomposition are widely employed to calculate the river flows; however, the influence of the application of diverse decomposing approaches on assessing correctness is inadequately compared and examined. This study investigates the certainty of monthly streamflow by applying ANNs including feed forward back propagation neural network and radial basis function neural network (RBFNN) models integrated with discrete wavelet transform (DWT), at Jinsha River basin in the upper reaches of Yangtze River of China. The effect of the noise factor of the decomposed time series on the prediction correctness has also been argued in this paper. Data have been analyzed by comparing the simulation outputs of the models with the correlation coefficient (R) root mean square errors, mean absolute errors, mean absolute percentage error and Nash–Sutcliffe Efficiency. Results show that time series decomposition technique DWT contributes in improving the accuracy of streamflow prediction, as compared to single ANN’s. The detailed comparative analysis showed that the RBFNN integrated with DWT has better forecasting capabilities as compared to other developed models. Moreover, for high-precision streamflow prediction, the high-frequency section of the original time series is very crucial, which is understandable in flood season.

Wavelet-linear genetic programming: A new approach for modeling monthly streamflow

The streamflows are important and effective factors in stream ecosystems and its accurate prediction is an essential and important issue in water resources and environmental engineering systems. A hybrid wavelet-linear genetic programming (WLGP) model, which includes a discrete wavelet transform (DWT) and a linear genetic programming (LGP) to predict the monthly streamflow (Q) in two gauging stations, Pataveh and Shahmokhtar, on the Beshar River at the Yasuj, Iran were used in this study. In the proposed WLGP model, the wavelet analysis was linked to the LGP model where the original time series of streamflow were decomposed into the sub-time series comprising wavelet coefficients. The results were compared with the single LGP, artificial neural network (ANN), a hybrid wavelet-ANN (WANN) and Multi Linear Regression (MLR) models. The comparisons were done by some of the commonly utilized relevant physical statistics. The Nash coefficients (E) were found as 0.877 and 0.817 for the WLGP model, for the Pataveh and Shahmokhtar stations, respectively. The comparison of the results showed that the WLGP model could significantly increase the streamflow prediction accuracy in both stations. Since, the results demonstrate a closer approximation of the peak streamflow values by the WLGP model, this model could be utilized for the simulation of cumulative streamflow data prediction in one month ahead.

Streamflow estimation by support vector machine coupled with different methods of time series decomposition in the upper reaches of Yangtze River, China

Machine learning models combined with time series decomposition are widely employed to estimate streamflow, yet the effect of the utilization of different decomposing methods on estimating accuracy is inadequately investigated and compared. In this paper, the main objective is to research the predictability of monthly streamflow using support vector machine model coupled with discrete wavelet transform (DWT) and empirical mode decomposition (EMD). The influence of the noise component of the decomposed time series on the forecast accuracy is also discussed here. Performance is evaluated through an application on Jinsha River, which is located in the upper reaches of Yangtze River in China. Results indicate that both time series decomposition techniques EMD and DWT contribute to improving the accuracy of streamflow prediction, and deeper comparative analysis shows models coupled with DWT have better prediction capabilities than models coupled with EMD. Furthermore, the high frequency component of the original series is indispensable for high-precision streamflow prediction, which is obvious in flood season.

Utilizing satellite precipitation estimates for streamflow forecasting via adjustment of mean field bias in precipitation data and assimilation of streamflow observations

Summary This study explores mitigating bias in satellite quantitative precipitation estimates (SQPE) and improving hydrologic predictions at ungauged locations via adjustment of the mean field bias (MFB) in SQPE and data assimilation (DA) of streamflow observations in a distributed hydrologic model. In this study, a variational procedure is used to adjust MFB in Climate Prediction Center MORPHing (CMORPH) SQPE and assimilate streamflow observations at the outlet of Elk River Basin in Missouri into the distributed Sacramento Soil Moisture Accounting (SAC-SMA) and kinematic wave routing models. The benefits of assimilation are assessed by comparing the streamflow predictions with or without DA at both the outlet and an upstream location, and by comparing the soil moisture grids forced by CMORPH SQPE against those forced by higher-quality multisensor quantitative precipitation estimates (MQPE) from National Weather Service. Special attention is given to the dependence of the efficacy of DA on the quality and latency of the SQPE, and the impact of dynamic correction of MFB in the SQPE via DA. The results show that adjusting MFB in CMORPH SQPE in addition to assimilating outlet flow reduces 66% of the bias in the CMORPH SQPE analysis and the RMSE of 12-h streamflow predictions by 81% at the outlet and 34–62% at interior locations of the catchment. Compared to applying a temporally invariant MFB for the entire storm, the DA-based, dynamic MFB correction reduces the RMSE of 6-h streamflow prediction by 63% at the outlet and 39–69% at interior locations. It is also shown that the accuracy of streamflow prediction deteriorates if the delineation of the precipitation area by CMORPH SQPE is significantly different, as measured by the Hausdorff distance, from that by MQPE. When compared with adjusting MFB in the CMORPH SQPE over the entire assimilation window, adjusting the MFB for all but the latest 18 h (i.e., the latency of CMORPH SQPE) within the assimilation window reduces the mean square error (MSE)-based skill score of streamflow predictions at the outlet by up to 0.08 and at interior locations by up to 0.13.

수문곡선의 감수부 특성을 고려한 기저유출 산정

수문곡선의 감수부는 기저유출의 특성을 반영하기 때문에 강우유출 모형과 기저유출분리법을 이용한 기저유출 산정과정에서 감수부의 특성을 고려해야한다. 따라서 본 연구는 감수특성을 고려하여 Soil and Water Assessment Tool(SWAT)의 보정에서 유량예측의 정확성을 높이고, 보정된 SWAT으로부터 예측된 유량으로부터 기저유출을 분리하고자 하였다. 이를 위하여 RECESS으로부터 산정된 alpha factor와 11개의 다른 매개변수를 자동보정모듈에 적용한 시나리오 (S1)와 SWAT의 매개변수인 alpha factor를 포함한 12개의 매개변수를 자동보정모듈에 적용한 시나리오 (S2)에 대해 SWAT을 이용해 유량 모의를 하였다. 또한, 두 시나리오에 대해 SWAT으로 예측된 유량을 Web-based Hydrograph Analysis Tool (WHAT)을 적용하여 기저유출을 산정하였다. 보정 결과는 유량에 대한 두 시나리오의 Nash-Sutcliffe Efficiency (NSE) 값들 사이에 큰 차이는 보이지 않았으나 기저유출의 경우 S1에 대한 NSE는 0.777이고, S2의 NSE 결과는 0.844로 다소 큰 차이를 보였다. 연평균 유량의 분포의 정량적 비교를 위한 관측유량과 상대오차를 산정하였으며 S1에 대하여 20.78%, S2에 대하여 6.59%의 상대오차를 보였다. 본 연구는 모형을 이용하여 예측된 유량으로부터 기저유출을 산정하는데 있어 감수부 특성의 중요성을 보여주었다. Recession of hydrograph gives a significant contribution to estimation of baseflow using rainfall-runoff models and baseflow separation methods, because recession affects baseflow. This study attempted to enhance the accuracy of streamflow predictions using a Soil and Water Assessment Tool (SWAT) model and to separate baseflow from the predicted streamflow. For this, this study used two scenarios: 1) to calibrate eleven parameters using an auto-calibration tool with the alpha factor obtained from RECESS (S1); and 2) to calibrate twelve SWAT parameters including alpha factor (one of SWAT parameters) using an auto-calibration tool (S2). Then, baseflow spearation from the predicted streamflow was conducted by using Web-based Hydrograph Analysis Tool (WHAT). The results show that there is no significant difference between Nash-Sutcliffe efficiency (NSE) values of S1 and S2 for calibrations to streamflow. However, calibrations to baseflow showed that NSEs are 0.777 for S1 and 0.844 for S2, which means a significant difference. Quantitatively compared to the observed streamflow, relative errors were 20.78 % for S1 and 6.59 % for S2. Finally, this study showed the importance of recession in baseflow separated from the predicted streamflow using a rainfall-runoff model.

The effect of spatial rainfall variability on water balance modelling for south-eastern Australian catchments

Summary Spatial rainfall variability is considered an important factor affecting the accuracy of streamflow prediction. This study evaluated the effect of spatial rainfall variability on water balance modelling by conducting a series of virtual experiments. The three-layer Variable Infiltration Capacity model (VIC-3L) was applied using both uniform and spatially variable rainfall for 60 catchments in south-eastern Australia. The spatially variable rainfall was generated from the 0.05° gridded SILO daily rainfall with different degrees of variability. The VIC-3L model was calibrated against observed daily streamflow using gridded SILO daily rainfall to generate reference parameter values. Then the model was applied using the generated spatially variable rainfall and reference parameter values to produce virtual water balance components associated with different spatially variable rainfall. Differences between the lumped and distributed modelling (i.e. virtual experiments) represent the effects of spatial rainfall variability on water balance modelling. The results showed that spatial rainfall variability interacts with catchments characteristics to influence hydrological processes and the effects are not uniform on different water balance components. For a given rainfall total, ignoring spatial rainfall variability will result in underestimation of the total streamflow volume and overestimation of evapotranspiration. Effect of spatial rainfall variability on water balance modelling is more pronounced in catchments with larger rainfall variability. In most cases, information on spatial rainfall variability will help to improve accuracy of water balance modelling. However, in some cases, lumped water balance modelling (i.e. ignoring spatial rainfall variability) may outperform distributed modelling due to catchment filtering effect. The results from this study suggest that more accurate information of rainfall spatial distribution in a catchment can help to improve water balance modelling.

Advancing monthly streamflow prediction accuracy of CART models using ensemble learning paradigms

► We use classification and regression trees for streamflow forecasting, first. ► We employ the support vector regression model as the benchmark model. ► We incorporate the bagging and boosting to advance the prediction accuracy, secondly. ► The results of ensembles are nearly same and superior to the single models. ► Bagged model performs slightly better than the boosted model in the testing phase. Streamflow forecasting is one of the most important steps in the water resources planning and management. Ensemble techniques such as bagging, boosting and stacking have gained popularity in hydrological forecasting in the recent years. The study investigates the potential usage of two ensemble learning paradigms (i.e., bagging; stochastic gradient boosting) in building classification and regression trees (CARTs) ensembles to advance the streamflow prediction accuracy. The study, initially, investigates the use of classification and regression trees for monthly streamflow forecasting and employs a support vector regression (SVR) model as the benchmark model. The analytic results indicate that CART outperforms SVR in both training and testing phases. Although the obtained results of CART model in training phase are considerable, it is not in testing phase. Thus, to optimize the prediction accuracy of CART for monthly streamflow forecasting, we incorporate bagging and stochastic gradient boosting which are rooted in same philosophy, advancing the prediction accuracy of weak learners. Comparing with the results of bagged regression trees (BRTs) and stochastic gradient boosted regression trees (GBRTs) models possess satisfactory monthly streamflow forecasting performance than CART and SVR models. Overall, it is found that ensemble learning paradigms can remarkably advance the prediction accuracy of CART models in monthly streamflow forecasting.

Suitability of the TRMM satellite rainfalls in driving a distributed hydrological model for water balance computations in Xinjiang catchment, Poyang lake basin

Summary Spatial rainfall is a key input to distributed hydrological models, and its precisions heavily affect the accuracy of stream flow predictions from a hydrological model. Traditional interpolation techniques which obtain the spatial rainfall distribution from rain gauge data have some limitations caused by data scarcity and bad quality, especially in developing countries or remote locations. Satellite-based precipitation products are expected to offer an alternative to ground-based rainfall estimates in the present and the foreseeable future. For this purpose, the quality and usefulness of satellite-based precipitation products need to be evaluated. The present study compares the difference of Tropical Rainfall Measuring Mission (TRMM) rainfall with rain gauges data at different time scales and evaluates the usefulness of the TRMM rainfall for hydrological processes simulation and water balance analysis at the Xinjiang catchment, located in the lower reaches of the Yangtze River in China. The results show at daily time step TRMM rainfall data are better at determining rain occurrence and mean values than at determining the rainfall extremes, and larger difference exists for the maximal daily and maximal 5-day rainfalls. At monthly time scale, good linear relationships between TRMM rainfall and rain gauges rainfall data are received with the determination coefficients (R2) varying between 0.81 and 0.89 for the individual stations and 0.88 for areal average rainfall data, respectively. But the slope of regression line ranges between 0.74 for Yingtan and 0.94 for Yushan, indicating that the TRMM satellite is inclined to underestimate the monthly rainfall in this area. The simulation of daily hydrological processes shows that the Water Flow Model for Lake Catchment (WATLAC) model using conventional rain gauge data produces an overall good fit, but the simulation results using TRMM rainfall data are discontented. The evaluation results imply that the TRMM rainfall data are unsuited for daily stream flow simulation in this study area with desired precisions. However, good performance can be received using TRMM rainfall data for monthly stream flow simulations. The comparison of the simulated annual water balance components shows that the different rainfall data sources can change the volume value and proportion of water balance components to some extent, but it generally meets the need of practical use.

Assessing the hydrologic performance of the EPA’s nonpoint source water quality assessment decision support tool using North American Land Data Assimilation System (NLDAS) products

The accuracy of streamflow predictions in the EPA’s BASINS (Better Assessment Science Integrating Point and Nonpoint Sources) decision support tool is affected by the sparse meteorological data contained in BASINS. The North American Land Data Assimilation System (NLDAS) data with high spatial and temporal resolutions provide an alternative to the NOAA National Climatic Data Center (NCDC)’s station data. This study assessed the improvement of streamflow prediction of the Hydrological Simulation Program-FORTRAN (HSPF) model contained within BASINS using the NLDAS 1/8 degree hourly precipitation and evapotranspiration estimates in seven watersheds of the Chesapeake Bay region. Our results demonstrated consistent improvements of daily streamflow predictions in five of the seven watersheds when NLDAS precipitation and evapotranspiration data was incorporated into BASINS. The improvement of using NLDAS data is significant when the watershed’s meteorological station is either far away or not in a similar climatic region. When the station is nearby, using NLDAS data produces similar results. The correlation coefficients of the analyses using NLDAS data were greater than 0.8, the Nash–Sutcliffe (NS) model fit efficiency greater than 0.6, and the error in the water balance was less than 5%. Our analyses also showed that the streamflow improvements were mainly contributed by NLDAS precipitation data and that the improvement from using NLDAS evapotranspiration data was not significant; partially due to the constraints of current BASINS-HSPF settings. However, NLDAS evapotranspiration data did improve the baseflow prediction. This study demonstrates NLDAS data has the potential to improve stream flow predictions, thus aid the water quality assessment in the EPA nonpoint water quality assessment decision tool.

Assessing the impacts of precipitation bias on distributed hydrologic model calibration and prediction accuracy

Physics-based distributed (PBD) hydrologic models predict runoff throughout a basin using the laws of conservation of mass and momentum, and benefit from more accurate and representative precipitation input. Vflo™ is a gridded distributed hydrologic model that predicts runoff and continuously updates soil moisture. As a participating model in the second Distributed Model Intercomparison Project (DMIP2), Vflo™ is applied to the Illinois and Blue River basins in Oklahoma. Model parameters are derived from geospatial data for initial setup, and then adjusted to reproduce the observed flow under continuous time-series simulations and on an event basis. Simulation results demonstrate that certain runoff events are governed by saturation excess processes, while in others, infiltration-rate excess processes dominate. Streamflow prediction accuracy is enhanced when multi-sensor precipitation estimates (MPE) are bias corrected through re-analysis of the MPE provided in the DMIP2 experiment, resulting in gauge-corrected precipitation estimates (GCPE). Model calibration identified a set of parameters that minimized objective functions for errors in runoff volume and instantaneous discharge. Simulated streamflow for the Blue and Illinois River basins, have Nash–Sutcliffe efficiency coefficients between 0.61 and 0.68, respectively, for the 1996–2002 period using GCPE. The streamflow prediction accuracy improves by 74% in terms of Nash–Sutcliffe efficiency when GCPE is used during the calibration period. Without model calibration, excellent agreement between hourly simulated and observed discharge is obtained for the Illinois, whereas in the Blue River, adjustment of parameters affecting both saturation and infiltration-rate excess processes were necessary. During the 1996–2002 period, GCPE input was more important than model calibration for the Blue River, while model calibration proved more important for the Illinois River. During the verification period (2002–2006), calibration was more important than GCPE input for the Illinois River, but both calibration and GCPE input improved prediction accuracy within the Blue River.

Optimization of a similarity measure for estimating ungauged streamflow

One approach to predicting streamflow in an ungauged catchment is to select an ensemble of hydrological models previously identified for similar gauged catchments, where the similarity is based on some combination of important physical catchment attributes. The focus of this paper is the identification of catchment attributes and optimization of a similarity measure to produce the best possible ungauged streamflow predictions given a data set and a conceptual model structure. As a case study, the SimHyd rainfall‐runoff model is applied to simulate monthly streamflow in 184 Australian catchments. Initial results show that none of 27 catchment attributes can be safely said to consistently give a better ensemble of models than random selection when used independently of other attributes. This is contrary to prior expectations and indicates the sparseness of information within our database of catchments, the importance in this case of prior knowledge for defining important attributes, and the potential importance of combining multiple attributes in order to usefully gauge similarity. Seven relatively independent attributes are then selected on the basis of prior knowledge. The weight with which each of these attributes contributes to the similarity measure is optimized to maximize streamflow prediction performance across a set of 95 catchments. The other 89 catchments are used to independently test the accuracy of streamflow predictions. Using the optimal set of weights led to marked improvement in the accuracy of predictions, showing that the method, while inferior to local calibration, is superior to alternative methods of model regionalization based on regression and spatial proximity. However, there is evidence of nonuniqueness in the optimal solution and the possibility that the attribute weights are somewhat dependent on the catchments used.

The effect of gauge sampling density on the accuracy of streamflow prediction for rural catchments

A streamflow model of a 4800 km2 rural catchment using distributed rain inputs and infiltration was set up in a computer. Rain inputs were provided in the form of half-hour accumulation maps of real storm events collected by radar over southern Quebec. Point sampling of the accumulation pattern simulated gauge returns. Interpolation of the ȁgaugeȁ data to complete coverage was accomplished using Thiessen polygons. Hydrographs were computed for the interpolated and radar accumulation fields as a simple function of distance from the catchment outflow point. Ten sampling densities were used, and for each density hydrographs were computed for a large number of repetitions with gauges redistributed randomly for each realization. Comparisons of estimates of total rainfall, total runoff, peak runoff and time to peak between the hydrographs resulting from the random sampling and interpolation, and that computed from the full resolution radar accumulation field were made. As with estimates of areal rainfall we find that gauge density has a very strong effect on the estimation accuracy of hydrograph parameters with the standard error generally falling off as a power law with increasing gauge density.