Mid‐Range Streamflow Forecasts Based on Climate Modeling – Statistical Correction and Evaluation1

Abstract: Mid‐range streamflow predictions are extremely important for managing water resources. The ability to provide mid‐range (three to six months) streamflow forecasts enables considerable improvements in water resources system operations. The skill and economic value of such forecasts are of great interest. In this research, output from a general circulation model (GCM) is used to generate hydrologic input for mid‐range streamflow forecasts. Statistical procedures including: (1) transformation, (2) correction, (3) observation of ensemble average, (4) improvement of forecast, and (5) forecast skill test are conducted to minimize the error associated with different spatial resolution between the large‐scale GCM and the finer‐scale hydrologic model and to improve forecast skills. The accuracy of a streamflow forecast generated using a hydrologic model forced with GCM output for the basin was evaluated by forecast skill scores associated with the set of streamflow forecast values in a categorical forecast. Despite the generally low forecast skill score exhibited by the climate forecasting approach, precipitation forecast skill clearly improves when a conditional forecast is performed during the East Asia summer monsoon, June through August.     

SUWANNEE RIVER LONG RANGE STREAMFLOW FORECASTS BASED ON SEASONAL CLIMATE PREDICTORS1

ABSTRACT: A study of the influence of climate variability on streamflow in the southeastern United States is presented. Using a methodology previously applied to watersheds in Australia and the United States, a long range streamflow forecast (0 to 9 months in advance) is developed. Persistence (i.e., the previous season's streamflow) and climate predictors of the previous season are used to forecast the following season's (winter and spring) streamflow of the Suwannee River located in northern Florida. The winter and spring streamflow is historically the most likely to have severe flood events due to large scale cyclonic (frontal) storms. Results of the analysis indicated that a strong El Nino‐Southern Oscillation (ENSO) signal exists at various lead times to the winter and spring streamflow of the Suwannee River. These results are based on the high correlation values of two commonly used measurements of ENSO strength, the Multivariate ENSO Index (MEI) and Sea Surface Temperature Range 1. Using the relationships developed between climate and streamflow, a continuous exceedance probability forecast was developed for two Suwannee River stations. The forecast system provided an improved forecast for ENSO years. The ability to predict above normal (flood) or below normal (drought) years can provide communities the necessary lead time to protect life, property, sensitive wetlands, and endangered and threatened species.     

LONG-RANGE STREAMFLOW FORCASTING USING NONPARAMETRIC REGRESSION1

ABSTRACT: A class of nonparametric procedures is developed for producing long-range streamflow forecasts. The forecasting procedures, which are based solely on daily streamflow data, utilize nonparametric regression to relate a forecast variable to a covariate variable. The forecast variable is a function of future streamflow and can take a wide variety of forms. The covariate variable is a function of antecedent streamflow. The forecasting procedures are quite flexible, both in terms of the duration of the forecast period and the types of forecast variables that can be considered. The procedures are used to develop long-term (1–4 months) forecasts of minimum daily flow of the Potomac River at Washington, D.C. This forecast information is an integral component of water management activities for the Washington, D.C. metropolitan area.     

Streamflow Forecasting Using Singular Value Decomposition and Support Vector Machine for the Upper Rio Grande River Basin

The current study improves streamflow forecast lead‐time by coupling climate information in a data‐driven modeling framework. The spatial–temporal correlation between streamflow and oceanic–atmospheric variability represented by sea surface temperature (SST), 500‐mbar geopotential height (Z₅₀₀), 500‐mbar specific humidity (SH₅₀₀), and 500‐mbar east–west wind (U₅₀₀) of the Pacific and the Atlantic Ocean is obtained through singular value decomposition (SVD). SVD significant regions are weighted using a nonparametric method and utilized as input in a support vector machine (SVM) framework. The Upper Rio Grande River Basin (URGRB) is selected to test the applicability of the proposed model for the period of 1965–2014. The April–August streamflow volume is forecasted using previous year climate variability, creating a lagged relationship of 1–13 months. SVD results showed the streamflow variability was better explained by SST and U₅₀₀ as compared to Z₅₀₀ and SH₅₀₀. The SVM model showed satisfactory forecasting ability with best results achieved using a one‐month lead to forecast the following four‐month period. Overall, the SVM results showed excellent predictive ability with average correlation coefficient of 0.89 and Nash–Sutcliffe efficiency of 0.79. This study contributes toward identifying new SVD significant regions and improving streamflow forecast lead‐time of the URGRB.     

Does Including Soil Moisture Observations Improve Operational Streamflow Forecasts in Snow‐Dominated Watersheds?

Changing climate and growing water demand are increasing the need for robust streamflow forecasts. Historically, operational streamflow forecasts made by the Natural Resources Conservation Service have relied on precipitation and snow water equivalent observations from Snow Telemetry (SNOTEL) sites. We investigate whether also including SNOTEL soil moisture observations improve April‐July streamflow volume forecast accuracy at 0, 1, 2, and 3‐month lead times at 12 watersheds in Utah and California. We found statistically significant improvement in 0 and 3‐month lead time accuracy in 8 of 12 watersheds and 10 of 12 watersheds for 1 and 2‐month lead times. Surprisingly, these improvements were insensitive to soil moisture metrics derived from soil physical properties. Forecasts were made with volumetric water content (VWC) averaged from October 1 to the forecast date. By including VWC at the 0‐month lead time the forecasts explained 7.3% more variability and increased the streamflow volume accuracy by 8.4% on average compared to standard forecasts that already explained an average 77% of the variability. At 1 to 3‐month lead times, the inclusion of soil moisture explained 12.3‐26.3% more variability than the standard forecast on average. Our findings indicate including soil moisture observations increased statistical streamflow forecast accuracy and thus, could potentially improve water supply reliability in regions affected by changing snowpacks.     

REAL-TIME QUALITY CONTROL OF STREAMFLOW DATA—A SIMULATION STUDY1

ABSTRACT: The problem of real-time quality control of streamflow data is addressed. Five methods are investigated via a Monte-Carlo simulation experiment based on streamflow data from Bird Creek basin in Oklahoma. The five methods include three deterministic approaches and two statistical approaches. The relative performance of the investigated methods is evaluated under hypothesized random mechanism generating isolated outliers. The deterministic method based on streamflow gradient analysis and the statistical method based on forecast residual analysis perform best in detecting such outliers.     

Climatic Fluctuations and Forecasting of Streamflow in the Lower Colorado River Basin1

Abstract:  Water‐resource managers need to forecast streamflow in the Lower Colorado River Basin to plan for water‐resource projects and to operate reservoirs for water supply. Statistical forecasts of streamflow based on historical records of streamflow can be useful, but statistical assumptions, such as stationarity of flows, need to be evaluated. This study evaluated the relation between climatic fluctuations and stationarity and developed regression equations to forecast streamflow by using climatic fluctuations as explanatory variables. Climatic fluctuations were represented by the Atlantic Multidecadal Oscillation (AMO), Pacific Decadal Oscillation (PDO), and Southern Oscillation Index (SOI). Historical streamflow within the 25‐ to 30‐year positive or negative phases of AMO or PDO was generally stationary. Monotonic trends in annual mean flows were tested at the 21 sites evaluated in this study; 76% of the sites had no significant trends within phases of AMO and 86% of the sites had no significant trends within phases of PDO. As climatic phases shifted in signs, however, many sites had nonstationary flows; 67% of the sites had significant changes in annual mean flow as AMO shifted in signs. The regression equations developed in this study to forecast streamflow incorporate these shifts in climate and streamflow, thus that source of nonstationarity is accounted for. The R² value of regression equations that forecast individual years of annual flow for the central part of the study area ranged from 0.28 to 0.49 and averaged 0.39. AMO was the most significant variable, and a combination of indices from both the Atlantic and Pacific Oceans explained much more variation in flows than only the Pacific Ocean indices. The average R² value for equations with PDO and SOI was 0.15.     

Towards Real‐Time Continental Scale Streamflow Simulation in Continuous and Discrete Space

The National Weather Service (NWS) forecasts floods at approximately 3,600 locations across the United States (U.S.). However, the river network, as defined by the 1:100,000 scale National Hydrography Dataset‐Plus (NHDPlus) dataset, consists of 2.7 million river segments. Through the National Flood Interoperability Experiment, a continental scale streamflow simulation and forecast system was implemented and continuously operated through the summer of 2015. This system leveraged the WRF‐Hydro framework, initialized on a 3‐km grid, the Routing Application for the Parallel Computation of Discharge river routing model, operating on the NHDPlus, and real‐time atmospheric forcing to continuously forecast streamflow. Although this system produced forecasts, this paper presents a study of the three‐month nowcast to demonstrate the capacity to seamlessly predict reach scale streamflow at the continental scale. In addition, this paper evaluates the impact of reservoirs, through a case study in Texas. Validation of the uncalibrated model using observed hourly streamflow at 5,701 U.S. Geological Survey gages shows 26% demonstrate PBias ≤ |25%|, 11% demonstrate Nash‐Sutcliffe Efficiency (NSE) ≥ 0.25, and 6% demonstrate both PBias ≤ |25%| and NSE ≥ 0.25. When evaluating the impact of reservoirs, the analysis shows when reservoirs are included, NSE ≥ 0.25 for 56% of the gages downstream while NSE ≥ 0.25 for 11% when they are not. The results presented here provide a benchmark for the evolving hydrology program within the NWS and supports their efforts to develop a reach scale flood forecasting system for the country.     

Using Paleo Reconstructions to Improve Streamflow Forecast Lead Time in the Western United States

In water stressed regions, water managers are exploring new horizons that would help in long‐range streamflow forecasts. Oceanic‐atmospheric oscillations have been shown to influence streamflow variability. In this study, long‐lead time streamflow forecasts are made using a multiclass kernel‐based data‐driven support vector machine (SVM) model. The extended streamflow records based on tree ring reconstructions were used to provide a longer time series data. Reconstructed data were used from 1658 to 1952 and the instrumental record was used from 1953 to 2007. Reconstructions for oceanic‐atmospheric oscillations included the El Niño‐Southern Oscillation, Pacific Decadal Oscillation, Atlantic Multidecadal Oscillation, and North Atlantic Oscillation. Streamflow forecasts using all four oscillations were made with one‐year to five‐year lead times for 21 gages in the western United States. This is the first study that uses both instrumental and reconstructed data of oscillations in SVM model to improve streamflow forecast lead time. SVM model was able to provide “satisfactory” to “very good” forecasts with one‐ to five‐year lead time for the selected gages. The use of all the oscillation indices helped in achieving better predictability compared to using individual oscillations. The SVM modeling results are better when compared with multiple linear regression model forecasts. The findings are statistical in nature and are expected to be useful for long‐term water resources planning and management.     

Development and Operational Testing of a Super‐Ensemble Artificial Intelligence Flood‐Forecast Model for a Pacific Northwest River

Coastal catchments in British Columbia, Canada, experience a complex mixture of rainfall‐ and snowmelt‐driven contributions to flood events. Few operational flood‐forecast models are available in the region. Here, we integrated a number of proven technologies in a novel way to produce a super‐ensemble forecast system for the Englishman River, a flood‐prone stream on Vancouver Island. This three‐day‐ahead modeling system utilizes up to 42 numerical weather prediction model outputs from the North American Ensemble Forecast System, combined with six artificial neural network‐based streamflow models representing various slightly different system conceptualizations, all of which were trained exclusively on historical high‐flow data. As such, the system combines relatively low model development times and costs with the generation of fully probabilistic forecasts reflecting uncertainty in the simulation of both atmospheric and terrestrial hydrologic dynamics. Results from operational testing by British Columbia's flood forecasting agency during the 2013‐2014 storm season suggest that the prediction system is operationally useful and robust.     

A Streamflow Forecasting Framework using Multiple Climate and Hydrological Models1

Abstract: Water resources planning and management efficacy is subject to capturing inherent uncertainties stemming from climatic and hydrological inputs and models. Streamflow forecasts, critical in reservoir operation and water allocation decision making, fundamentally contain uncertainties arising from assumed initial conditions, model structure, and modeled processes. Accounting for these propagating uncertainties remains a formidable challenge. Recent enhancements in climate forecasting skill and hydrological modeling serve as an impetus for further pursuing models and model combinations capable of delivering improved streamflow forecasts. However, little consideration has been given to methodologies that include coupling both multiple climate and multiple hydrological models, increasing the pool of streamflow forecast ensemble members and accounting for cumulative sources of uncertainty. The framework presented here proposes integration and offline coupling of global climate models (GCMs), multiple regional climate models, and numerous water balance models to improve streamflow forecasting through generation of ensemble forecasts. For demonstration purposes, the framework is imposed on the Jaguaribe basin in northeastern Brazil for a hindcast of 1974‐1996 monthly streamflow. The ECHAM 4.5 and the NCEP/MRF9 GCMs and regional models, including dynamical and statistical models, are integrated with the ABCD and Soil Moisture Accounting Procedure water balance models. Precipitation hindcasts from the GCMs are downscaled via the regional models and fed into the water balance models, producing streamflow hindcasts. Multi‐model ensemble combination techniques include pooling, linear regression weighting, and a kernel density estimator to evaluate streamflow hindcasts; the latter technique exhibits superior skill compared with any single coupled model ensemble hindcast.     

SPATIAL VARIATIONS OF WATER LOSS DURING DROUGH A CASE STUDY1

ABSTRACT: A regional water conservation system for drought management involves many uncertain factors. Water received from precipitation may stay on the ground surface, evaporate back into the atmosphere, or infiltrate into the ground. Reliable estimates of the amount of evapotranspiration and infiltration are not available for a large basin, especially during periods of drought. By applying a geographic information system, this study develops procedures to investigate spatial variations of unavailable water for given levels of drought severity. Levels of drought severity are defined by truncated values of monthly precipitation and daily streamflow to reflect levels of water availability. The greater the truncation level, the lower the precipitation or streamflow. Truncation levels of monthly precipitation are recorded in depth of water while those of daily streamflow are converted into monthly equivalent water depths. Truncation levels of precipitation and streamflow treated as regionalized variables are spatially interpolated by the unbiased minimum variance estimation. The interpolated results are vector values of precipitation and streamflow at a grid of points covering the studied basin. They are then converted into raster‐based values and expressed graphically. The image subtraction operation is used to subtract the image of streamflow from that of precipitation at their corresponding level of drought severity. It is done on a cell‐by‐cell basis resulting in new attribute values to form the spatial image representing a spatial distribution of potential water loss at a given level of drought severity.     

BENEFITS OF TRANSFERRING STREAMFLOW PRIORITY FROM AGRICULTURAL TO NON-AGRICULTURAL USE1

ABSTRACT: In Virginia, as in many states, priority to streamflow is held by riparian landowners who are predominantly agricultural users. The streamfiow may also have a high potential value to non-agricultural users who do not have riparian rights. The potential benefits of transferring streamfiow priority rights from agricultural to non-agricultural use were evaluated using simulation for an eastern Virginia watershed. Lowering irrigators' priority to streamflow reduced crop yields and irrigated returns in some years because of inadequate water supplies. However, the transfer of priorities increased the likelihood that the urban reservoir would be able to withdraw water from the stream without interruption. As a result, priority trades reduced the size of reservoir needed to meet a given water requirement by municipal users. The resulting savings in reservoir construction and maintenance costs more than offset the losses to irrigators. Net savings could be achieved even if the reservoir were required to release water periodically to maintain a minimum level of instream flow. The conclusion is that the state should encourage trading of access to streamflow in order to increase the use efficiency of streamfiows. Alternative means by which the state can facilitate water exchanges are discussed.     

AI-based runoff simulation based on remote sensing observations: A case study of two river basins in the United States and Canada

Data-driven techniques are used extensively for hydrologic time-series prediction. We created various data-driven models (DDMs) based on machine learning: long short-term memory (LSTM), support vector regression (SVR), extreme learning machines, and an artificial neural network with backpropagation, to define the optimal approach to predicting streamflow time series in the Carson River (California, USA) and Montmorency (Canada) catchments. The moderate resolution imaging spectroradiometer (MODIS) snow-coverage dataset was applied to improve the streamflow estimate. In addition to the DDMs, the conceptual snowmelt runoff model was applied to simulate and forecast daily streamflow. The four main predictor variables, namely snow-coverage (S-C), precipitation (P), maximum temperature (T_max), and minimum temperature (T_min), and their corresponding values for each river basin, were obtained from National Climatic Data Center and National Snow and Ice Data Center to develop the model. The most relevant predictor variable was chosen using the support vector machine-recursive feature elimination feature selection approach. The results show that incorporating the MODIS snow-coverage dataset improves the models' prediction accuracies in the snowmelt-dominated basin. SVR and LSTM exhibited the best performances (root mean square error = 8.63 and 9.80) using monthly and daily snowmelt time series, respectively. In summary, machine learning is a reliable method to forecast runoff as it can be employed in global climate forecasts that require high-volume data processing.     

Analyzing the Effects of Groundwater Pumping on an Urban Stream‐Aquifer System

When the cone of influence of a pumping well reaches a nearby river, the resulting hydraulic gradient can induce enhanced seepage of streamflow into the aquifer. The rate of seepage is often modeled using analytical solutions that are simple to apply but may not reproduce field data due to mathematical assumptions not being met in the field. Furthermore, the appropriateness of such models has not been investigated in detail due to difficulty in measuring streamflow loss in the field. In this study, a field experiment was conducted on a reach of the South Platte River near Denver, Colorado to estimate pumping‐induced streamflow loss. A network of stream gauges, monitoring wells, and in situ measurements was used to observe streamflow rates, groundwater levels, and temperature to assess if pumping wells have a significant impact on streamflow, and to compare observed streamflow depletion against analytical solutions. Data collected suggest that pumping wells have a noticeable impact on streamflow. The analytical solutions proved accurate if streamflow was low and constant but performed poorly if streamflow was high and variable. Therefore, for this reach, the use of analytical solutions to predict streamflow may only be appropriate under low‐flow, constant‐flow conditions. Methods and results can be used to guide other streamflow depletion studies and to inform cases of pumping‐induced streamflow depletion, particularly in regard to water rights.     

A High‐Resolution National‐Scale Hydrologic Forecast System from a Global Ensemble Land Surface Model

Warning systems with the ability to predict floods several days in advance have the potential to benefit tens of millions of people. Accordingly, large‐scale streamflow prediction systems such as the Advanced Hydrologic Prediction Service or the Global Flood Awareness System are limited to coarse resolutions. This article presents a method for routing global runoff ensemble forecasts and global historical runoff generated by the European Centre for Medium‐Range Weather Forecasts model using the Routing Application for Parallel computatIon of Discharge to produce high spatial resolution 15‐day stream forecasts, approximate recurrence intervals, and warning points at locations where streamflow is predicted to exceed the recurrence interval thresholds. The processing method involves distributing the computations using computer clusters to facilitate processing of large watersheds with high‐density stream networks. In addition, the Streamflow Prediction Tool web application was developed for visualizing analyzed results at both the regional level and at the reach level of high‐density stream networks. The application formed part of the base hydrologic forecasting service available to the National Flood Interoperability Experiment and can potentially transform the nation's forecast ability by incorporating ensemble predictions at the nearly 2.7 million reaches of the National Hydrography Plus Version 2 Dataset into the national forecasting system.     

FORECAST UNCERTAINTY IN WATER SUPPLY RESERVOIR OPERATION1

ABSTRACT: This research investigates the benefits of forecasting in water supply systems. Questions relating operational losses to forecast period and accuracy are addressed. Some simple available forecasting techniques are assessed for their accuracy and applicability. These issues are addressed through the use of a simulation model of the Cedar and South Fork Tolt Rivers, where the system is modeled as a single purpose reservoir supplying municipal and industrial water to the Seattle metropolitan area. The following conclusions were made for this system: (1) reservoir operation deteriorates markedly with the loss of forecast accuracy; (2) the optimal length of forecasting period is five months; (3) reservoir operation may be improved by as much as 88 percent if perfect predictive abilities are available; (4) the mean of the historic data is not recommended to predict future flows because Markov methods are always superior; and (5) lag-one autoregressive Markov schemes exhibit about a 9 percent improvement in operation over no forecasting.     

A STOCHASTIC DYNAMIC PROGRAMMING BASED APPROACH TO THE OPERATION OF A MULTI-RESERVOIR SYSTEM1

ABSTRACT: An heuristic iterative technique based upon stochastic dynamic programming is presented for the analysis of the operation of a three reservoir ‘Y’ shaped hydroelectric system. The technique is initiated using historical inflow data for the downstream reservoir. At each iteration the optimal policies for the downstream hydroelectric generating unit are used to provide relative weightings or targets for operation of upstream reservoirs. New input inflows to the downstream reservoir are then obtained by running the historical streamflow record through the optimal policies for the upstream reservoirs. These flows are then used to develop a new operating policy for the downstream reservoir and hence new targets for the upstream reservoirs. The process is continued until the operating policies for each reservoir provide the same overall system benefit for two successive iterations. Results obtained from the procedure are compared to the results obtained by historical operation of the system. The procedure is shown to develop operating policies which give benefits which are as close to the historical benefits as can be expected given the choice of the number of storage state variables.     

Exploring Persistence in Streamflow Forecasting

In this study, the authors explore three persistence approaches in streamflow forecasting motivated by the need for forecasting model skill evaluation. The authors use streamflow observations with 15 min resolution from the year 2008 to 2017 at 140 United States Geological Survey streamflow gauges monitoring the streams and rivers over the State of Iowa. The spatial scale of the basins ranges from about 7 to 37,000 km². The study explores three approaches: simple persistence, gradient persistence, and anomaly persistence. The study shows that persistence forecasts skill has strong dependence on basin scales and weaker but non‐negligible dependence on geometric properties of the river network for a given basin. Among the three approaches explored, anomaly persistence shows highest skill especially for small basins, under about 500 km². The anomaly persistence can serve as a benchmark for model evaluations considering the effect of basin scales and geometric properties of river network of the basin. This study further reiterates that persistence forecasts are hard‐to‐beat methods for larger basin scales at short to medium forecast range.     

ANN OUTPUT UPDATING OF LUMPED CONCEPTUAL RAINFALL/RUNOFF FORECASTING MODELS1

ABSTRACT: Artificial neural networks (ANNs) are tested for the output updating of one‐day‐ahead and three‐day‐ahead streamflow forecasts derived from three lumped conceptual rainfall/runoff (R‐R) models: the GR4J, the IHAC, and the TOPMO. ANN output updating proved superior to a parameter updating scheme and to the ‘simple’ output updating scheme, which always replicates the last observed forecast error. In fact, ANN output updating was able to compensate for large differences in the initial performance of the three tested lumped conceptual R‐R models, which the other tested updating approaches were not able to achieve. This is done mainly by incorporating input vectors usually exploited for ANN R‐R modeling such as previous rainfall and streamflow observations, in addition to the previous observed error. For one‐day‐ahead forecasts, the performance of all three lumped conceptual R‐R models, used in conjunction with ANN output updating, was equivalent to that of the ANN R‐R model. For three‐day‐ahead forecasts, the performance of the ANN‐output‐updated conceptual models was even superior to that of the ANN R‐R model, revealing that the conceptual models are probably performing some tasks that the ANN R‐R model cannot map. However, further testing is needed to substantiate the last statement.