Ensemble and Probabilistic Estimation and Forecasting of Precipitation for Hydrologic Prediction I
Presiding: D Seo, NOAA/NWS/Office of Hydrologic Development/Hydrology Laboratory; A Bradley, IIHR-Hydroscience and Engineering, University of Iowa
H43E-01 13:30h
Three-tier operational probabilistic precipitation and hydrological forecasting of large-scale un-gauged river basins: Developing a basis for strategic and tactical decisions for water management, agricultural planning
The problem is that a user or user community must make deterministic (yes/no) decisions on multiple time scales about environmental systems containing considerable uncertainty. Such decisions range from strategic to tactical. A useful product takes into account forecasts of the probabilities of a state of the environment (e.g., rain rate, river discharge, flood potential and so on) and quantitative or qualitative information from the user about the consequences of occurrence of a particular environmental state. Considered in tandem, optimal hedging and risk analysis can be undertaken. We use Bangladesh as an example of the three-tiered forecast system and its application to real problems. Because of lack of Ganges and Brahmaputra river flow information from India upriver of the borders of the Bangaldesh, we have had to treat the Ganges and Brahmaputra river catchment areas as essentially ungauged. Through the development of hydrological techniques in conjunction with ensemble forecasts from ECMWF, river discharge forecasts are now supplied to Government of Bangladesh authorities in real time throughout the summer period. In addition, as a biproduct of these efforts, we also produce regional precipitation forecasts for Bangladesh and a number of regions within India. Forecasts of river discharge on 1-6 month periods (using the ECMWF couple climate model and hydrological models and statistical methods) are provided every month starting in April. Last year, we were able to forecast the July floods quite well some 3 months in advance. Using Bayesian empirical scheme, we make pentad 20-day forecasts every 5 days of regional rainfall and river discharge. Finally, we use the ECMWF ensembles of forecasts (51 predictions per day) to issue 1-10 day probabilistic forecasts of river discharge and precipitation. River discharge is made with the additional use of a suite of hydrological models. In addition, we provide threshold probability forecasts on the 1-10 day time scale of significant occurrences such as flood danger level. Finally, these forecasts are cast within an easily understood "user metric" which combines the probabilistic forecast with information of impact from the user community.
H43E-02 13:45h
Probabilistic Quantitative Precipitation Forecasting Based on Ensemble Reforecasts
We explore how probabilistic quantitative precipitation forecasts (PQPFs) can be improved through judicious calibration of the current ensemble weather forecast using a long training data set of retrospective forecasts (reforecasts). The Climate Diagnostics Center of NOAA has developed a 25-year retrospective database of 2-week, 15-member ensemble weather forecasts, all run with the same weather forecast model throughout the 25 years. This provides a long-term, stable database of weather forecasts from which systematic errors can be diagnosed and corrected.
H43E-03 14:00h
Dependence of Global Model Ensemble Precipitation Forecast Skill on Space and Time Scales
It is well known that some of the skill in precipitation forecasts comes from aggregation, both in space and time. Global model ensemble precipitation forecasts provide useful information on large scale precipitation patterns, but they have to be downscaled to match the input requirements of hydrologic forecast models. Reconstruction of detailed space-time precipitation variability is hence needed, but there is no one particular space-time scale combination that fits all hydrologic models operating at different scales. In this work, we investigate the spatial scale-dependency of the correlation between ensemble mean forecasts and the observed data for a range of space-time scales as a function of forecast lead time. The objective of the study is to develop statistical procedures for down/re-scaling precipitation fields that preserve forecast skill at all space and time scales of interest in support of the NOAA/National Weather Service's (NWS) ensemble hydrologic prediction. The study area is the continental U.S.A. The analysis is done using the daily global model ensemble precipitation reforecasts produced by the NOAA/Climate Diagnostic Center from a frozen version of the Global Forecast System of the NOAA/NWS/National Centers for Environmental Prediction. The 2-week ensemble reforecasts are archived on a 2.5 degree grid and are available for the 1979 - 2004 period. The observed precipitation data used for the analysis comes from the NOAA/National Climatic Data Center.
H43E-04 14:15h
Precipitation Ensembles from Single-Value Forecasts for Hydrological Ensemble Forecasting
An ensemble pre-processor was developed to produce short-term precipitation ensembles using operational single-value forecasts. The methodology attempts to quantify the uncertainty in the single-value forecast and to capture the skill therein. These precipitation ensemble forecasts could be then ingested in the NOAA/National Weather Service (NWS) Ensemble Streamflow Prediction (ESP) system to produce probabilistic hydrological forecasts that reflect the uncertainty in forecast precipitation. The procedure constructs the joint distribution of forecast and observed precipitation from historical pairs of forecast and observed values. The probability distribution function of the future events that may occur given a particular single-value forecast is then the conditional distribution of observed precipitation given the forecast. To generate individual ensemble members for each lead time and each location, the historical observed values are replaced with values sampled from the conditional distribution given the single-value forecast. The replacement procedure matches the ranks of historical and rescaled values to preserve the space-time properties of observed precipitation in the ensemble traces. Currently, the ensemble pre-processor is being tested and evaluated at four NOAA/NWS River Forecast Centers (RFCs) in the U.S.A. In this contribution, we present the results thus far from the field and retrospective evaluations, and key science issues that must be addressed toward national operational implementation.
H43E-05 14:30h
Impact of Spatial Interpolation Methods for Precipitation on Ensemble Streamflow Simulation From Watershed Models
Watershed models are used for simulating basin streamflows based on spatially sparse precipitation and temperature observations. The sparse observations are typically interpolated on a regular grid or a subbasin as inputs to the hydrologic models. Given the paucity in observations and nonhomogenous nature of the precipitation process, differences in interpolation methods can potentially impact the simulated streamflow. Of course, hydrologic model parameter uncertainty also contribute to the errors, but in this paper we focus on the uncertainty due to interpolation methods. To this end, first we developed a two-step process in which the precipitation occurrence is first generated via a logistic regression model, and the amounts are then estimated using a Multiple Linear Regression (MLR) and Locally Weighted Polynomial Regression (LWP). The two-step approach is shown to capture the spatial variability of precipitation effectively than other competing traditional methods. Secondly, interpolated precipitation estimates are input into the watershed model, Precipitation Runoff Modeling System (PRMS) to estimate daily and consequently, monthly and seasonal streamflows. Streamflow estimates from PRMS are obtained for three methods of precipitation interpolation, MLR, LWP and the currently used method in PRMS, Climatological MLR (CMLR). Streamflows are compared on a variety of attributes. We find that the MLR and LWP methods perform much better in simulating the streamflows compared to CMLR. Ensembles of precipitation from the two methods (MLR and LWP) coupled with the logistic regression for precipitation occurrence, are generated to subsequently generate ensembles of streamflows from the watershed model. This approach captures the input uncertainty.
http://ceae.colorado.edu/~yhwang/AGUmay
H43E-06 14:45h
Radar Based Probabilistic Quantitative Precipitation Estimation: First Results of Large Sample Data Analysis
Large uncertainties in the operational precipitation estimates produced by the U.S. national network of WSR-88D radars are well-acknowledged. However, quantitative information about these uncertainties is not operationally available. In an effort to fill this gap, the U.S. National Weather Service (NWS) is supporting the development of a probabilistic approach to the radar precipitation estimation. The probabilistic quantitative precipitation estimation (PQPE) methodology that was selected for this development is based on the empirically-based modeling of the functional-statistical error structure in the operational WSR-88D precipitation products under different conditions. Our first goal is to deliver a realistic parameterization of the probabilistic error model describing its dependences on the radar-estimated precipitation value, distance from the radar, season, spatiotemporal averaging scale, and the setup of the precipitation processing system (PPS). In the long-term perspective, when large samples of relevant data are available, we will extend the model to include the dependences on different types of precipitation estimates (e.g. polarimeteric and multi-sensor), geographic locations and climatic regimes. At this stage of the PQPE project, we organized a 6-year-long sample of the Level II data from the Oklahoma City radar station (KTLX), and processed it with the Built 4 of the PPS that is currently used in the NWS operations. This first set of operational products was generated with the standard setup of the PPS parameters. The radar estimates are completed with the corresponding raingauge data from the Oklahoma Mesonet, the ARS Little Washita Micronet and the EVAC PicoNet covering different spatial scales. The raingauge data are used as a ground reference (GR) to estimate the required uncertainty characteristics in the radar precipitation products. In this presentation, we describe the first results of the large-sample uncertainty analysis of the products. Based on the available surface temperatures, the data-sample is divided into three seasons: warm, intermediate and cold. For each of these sub-samples we estimate several conditional statistics using a nonparametric functional estimation technique. We discuss the possible parametric models that can reproduce the estimated dependences with satisfactory accuracy.