Abstract

When hydrological models are used for probabilistic streamflow forecasting in the Ensemble Streamflow Prediction (ESP) framework, the deterministic components of the approach can lead to errors in the estimation of forecast uncertainty, as represented by the spread of the forecast ensemble. One avenue for correcting the resulting forecast reliability errors is to calibrate the streamflow forecast ensemble to match observed error characteristics. This paper outlines and evaluates a method for forecast calibration as applied to seasonal streamflow prediction. The approach uses the correlation of forecast ensemble means with observations to generate a conditional forecast mean and spread that lie between the climatological mean and spread (when the forecast has no skill) and the raw forecast mean with zero spread (when the forecast is perfect). Retrospective forecasts of summer period runoff in the Feather River basin, California, are used to demonstrate that the approach improves upon the performance of traditional ESP forecasts by reducing errors in forecast mean and improving spread estimates, thereby increasing forecast reliability and skill.

Streamflow forecasting is critical to water resources management in the western United States. Yet, despite the passage of almost 50 years since the development of the first computerized hydrologic simulation models and over 30 years since the development of hydrologic ensemble forecast methods, the prevalent method used for forecasting seasonal streamflow in the western United States remains the regression of spring and summer streamflow volume on spring snowpack and/or the previous winter's accumulated precipitation. A recent retrospective analysis have shown that the skill of the regression-based forecasts have not improved in the last 40 years, despite large investments in science and technology related to the monitoring and assessment of the land surface and in climate forecasting. We describe an experimental streamflow forecast system for the western United States that applies a modern macroscale land surface model (akin to those now used in numerical weather prediction and climate models) to capture hydrologic states (soil moisture and snow) at the time of forecast, incorporates data assimilation methods to improve estimates of initial state, and uses a range of climate prediction ensembles to produce ensemble forecasts of streamflow and associated hydrologic states for lead times of up to one year. The forecast system is intended to be a real-time test bed for evaluating new seasonal streamflow forecast methods. Experience with the forecast system is illustrated using results from the 2004/05 forecast season, in which an evolving drought in the Pacific Northwest diverged strikingly from extreme snow accumulations to the south. We also discuss how the forecast system relates to ongoing changes in seasonal streamflow forecast methods in the two U.S. operational agencies that have major responsibility for seasonal streamflow forecasts in the western United States.

Abstract

Hydrologic model calibration is usually a central element of streamflow forecasting based on the ensemble streamflow prediction (ESP) method. Evaluation measures of forecast errors such as root-mean-square error (RMSE) are heavily influenced by bias, which in turn is readily reduced by calibration. On the other hand, bias can also be reduced by postprocessing (e.g., “training” bias correction schemes based on retrospective simulation error statistics). This observation invites the question: How much is forecast error reduced by calibration, beyond what can be accomplished by postprocessing to remove bias? The authors address this question through retrospective evaluation of forecast errors at eight streamflow forecast locations distributed across the western United States. Forecast periods of length ranging from 1 to 6 months are investigated, for forecasts initiated from 1 December to 1 June, which span the period when most runoff occurs from snowmelt-dominated western U.S. rivers. ESP forecast errors are evaluated both for uncalibrated forecasts to which a percentile mapping bias correction approach is applied, and for forecasts from an objectively calibrated model without explicit bias correction. Using the coefficient of prediction (C_p ), which essentially is a measure of the fraction of variance explained by the forecast, the authors find that the reduction in forecast error as measured by C_p that is achieved by bias correction alone is nearly as great as that resulting from hydrologic model calibration.

Abstract

Accurate precipitation data are critical for hydrologic prediction, yet outside the developed world in situ networks are so sparse as to make alternative methods of precipitation estimation essential. Several such alternative precipitation products that would be adequate to drive hydrologic prediction models at regional and global scales are evaluated. As a benchmark, a gridded station-based dataset is used, which is compared with the global 40-yr ECMWF Re-Analysis (ERA-40), and a satellite-based dataset [i.e., the Global Precipitation Climatology Project One-Degree Daily (GPCP 1DD)]. Each dataset, with a common set of other meteorological forcings aside from precipitation, was used to force the Variable Infiltration Capacity (VIC) macroscale hydrology model globally for the 1997–99 period for which the three datasets overlapped. The three precipitation datasets and simulated hydrological variables (i.e., soil moisture, runoff, evapotranspiration, and snow water equivalent) are compared in terms of the implied water balances of the continents, and for prediction of streamflow for nine large river basins. The evaluations are in general agreement with previous but more local evaluations of precipitation products and water balances: the precipitation datasets agree reasonably on the seasonality but less on monthly anomalies. Furthermore, the largest differences in precipitation are in mountainous regions and regions where in situ networks are sparse (such as Africa). Derived runoff is highly sensitive to differences in precipitation forcings. At a global level, all three simulations result in water budgets that are within the range of other water balance climatologies. Although uncertainties in the three datasets preclude an evaluation of which one has the lowest errors, overall ERA-40 is preferred because of its agreement with the station-based dataset in locations where the station density is high, its periodic availability, and its temporal resolution.

Abstract

A hydrometric network design approach is developed for enhancing statistical seasonal streamflow forecasts. The approach employs gridded, model-simulated water balance variables as predictors in equations generated via principal components regression in order to identify locations for additional observations that most improve forecast skill. The approach is applied toward the expansion of the Natural Resources Conservation Service (NRCS) Snowpack Telemetry (SNOTEL) network in 24 western U.S. basins using two forecasting scenarios: one that assumes the currently standard predictors of snow water equivalent and water year-to-date precipitation and one that considers soil moisture as an additional predictor variable. Resulting improvements are spatially and temporally analyzed, attributed to dominant predictor contributions, and evaluated in the context of operational NRCS forecasts, ensemble-based National Weather Service (NWS) forecasts, and historical as-issued NRCS/NWS coordinated forecasts. Findings indicate that, except for basins with sparse existing networks, substantial improvements in forecast skill are only possible through the addition of soil moisture variables. Furthermore, locations identified as optimal for soil moisture sensor installation are primarily found in regions of low to mid elevation, in contrast to the higher elevations where SNOTEL stations are traditionally situated. The study corroborates prior research while demonstrating that soil moisture data can explicitly improve operational water supply forecasts (particularly during the accumulation season), that statistical forecasts are comparable in skill to ensemble-based forecasts, and that simulated hydrologic data can be combined with observations to improve statistical forecasts. The approach can be generalized to other settings and applications involving the use of point observations for statistical prediction models.

Abstract

Operational hydrologic models are typically calibrated using meteorological inputs derived from retrospective station data that are commonly not available in real time. Inconsistencies between the calibration and (generally sparser) real-time station datasets can be a source of bias, which can be addressed by expressing real-time hydrological model forcings (primarily precipitation) as percentiles for a set of index stations that report both in real time and during the retrospective calibration period, and by using the real-time percentiles to create adjusted precipitation forcings. Although hydrological model precipitation forcings typically are required at time steps of one day or shorter, percentiles can be calculated for longer averaging periods to reduce the percentile estimation errors. The authors propose an index station percentile method (ISPM) to estimate precipitation at the models input time step using percentiles, relative to a climatological period, for a set of index stations that report in real time. In general, this approach is most appropriate to situations in which the spatial correlation of precipitation is high, such as cold season rainfall in the western United States. The authors evaluate the ISPM approach, including performance sensitivity to the choice of percentile estimation period length, using the Klamath River basin, Oregon, as a case study. Relative to orographically adjusted interpolation of the real-time index station values, ISPM gives better estimates of precipitation throughout the basin. The authors find that ISPM performs best for percentile estimation periods longer than 10 days, with diminishing returns for averaging periods longer than 30 days. They also evaluate the performance of ISPM for a reduced station scenario and find that performance is relatively stable, relative to the competing methods, as the number of real-time stations diminishes.

Abstract

Subseasonal to seasonal (S2S) climate forecasting has become a central component of climate services aimed at improving water management. In some cases, operational S2S climate predictions are translated into inputs for follow-on analyses or models, whereas the S2S predictions on their own may provide for qualitative situational awareness. At the spatial scales of water management, however, S2S climate forecasts often suffer from systematic biases, and low skill and reliability. This study assesses the potential to improve S2S forecast skill and salience for watershed applications through the use of postprocessing to harness skills in large-scale fields from the global climate model forecast outputs. To this end, the components-based technique—partial least squares regression (PLSR)—is used to improve the skill of biweekly temperature and precipitation forecasts from the Climate Forecast System version 2 (CFSv2). The PLSR method forms predictor components based on a cross-validated analysis of hindcasts from CFSv2 climate and land surface fields, and the results are benchmarked against raw CFSv2 forecasts, remapped to intermediate-scale watershed areas. We find that postprocessing affords marginal to moderate gains in skill in many watersheds, raising climate forecast skill above a usability threshold over the four seasons analyzed. In other locations, however, postprocessing fails to improve skill, particularly for extreme events, and can lead to unreliably narrow forecast ranges. This work presents evidence that the statistical postprocessing of climate forecast system outputs has potential to improve forecast skill, but that more thorough study of alternative approaches and predictors may be needed to achieve comprehensively positive outcomes.

Abstract

The concepts of model benchmarking, model agility, and large-sample hydrology are becoming more prevalent in hydrologic and land surface modeling. As modeling systems become more sophisticated, these concepts have the ability to help improve modeling capabilities and understanding. In this paper, their utility is demonstrated with an application of the physically based Variable Infiltration Capacity model (VIC). The authors implement VIC for a sample of 531 basins across the contiguous United States, incrementally increase model agility, and perform comparisons to a benchmark. The use of a large-sample set allows for statistically robust comparisons and subcategorization across hydroclimate conditions. Our benchmark is a calibrated, time-stepping, conceptual hydrologic model. This model is constrained by physical relationships such as the water balance, and it complements purely statistical benchmarks due to the increased physical realism and permits physically motivated benchmarking using metrics that relate one variable to another (e.g., runoff ratio). The authors find that increasing model agility along the parameter dimension, as measured by the number of model parameters available for calibration, does increase model performance for calibration and validation periods relative to less agile implementations. However, as agility increases, transferability decreases, even for a complex model such as VIC. The benchmark outperforms VIC in even the most agile case when evaluated across the entire basin set. However, VIC meets or exceeds benchmark performance in basins with high runoff ratios (greater than ~0.8), highlighting the ability of large-sample comparative hydrology to identify hydroclimatic performance variations.

Abstract

Most datasets of surface meteorology are deterministic, yet many applications using these datasets require or can benefit from uncertainty estimates in meteorological fields. Motivated by this gap, we evaluated the use of a spatial regression method to estimate the uncertainty in precipitation and temperature fields of existing deterministic gridded meteorological datasets. Taking the widely used North American Land Data Assimilation System 2 (NLDAS-2) precipitation and temperature dataset as an example, we used the deterministic NLDAS-2 values to generate ensemble estimates for daily precipitation, mean temperature, and the diurnal temperature range. Our method is a form of ensemble dressing. Nine variations were tested to assess the impacts of sampling density on the estimates of the mean and uncertainty, and one strategy was selected to generate 100 ensemble members at 1/8° and daily resolution for the period 1979–2019, termed as the Ensemble Dressing of NLDAS-2 (EDN2). Compared with an independent station-based ensemble dataset, the ensemble dressing method produces reasonable uncertainty patterns for precipitation and underestimates uncertainty for temperature. For precipitation, the uncertainty increases with the increase in daily accumulation. For temperature, the uncertainty is relatively small in the warm season and large in the cold season. This ensemble dressing method is applicable to other deterministic gridded meteorological datasets. The generated spatiotemporally varying uncertainty information could support applications such as land surface and hydrologic modeling, data assimilation, and forecasting, especially where application models are tied to a specific meteorological dataset.