H32C-01 INVITED
Multi-Model Ensembling for Seasonal-to-Interannual Climate Forecasting at the IRI
Ultimately one wants a climate forecast that is as sharp as possible (i.e. high probabilities for a specific outcome) with probabilities that are as reliable as possible (i.e. true to the observed frequency of that outcome over time). Currently, there are many approaches to combining probabilistic predictions from GCMs. Many, but not all, consider some measure of model skill in determining the regional weighting given to a particular GCM. With reasonable sample sizes, most multi-model ensembling approaches yield an overall improvement on skill relative to the performance of the individual models. Occasionally, the multi-model ensemble even outperforms the best model over a particular region. The simplest approach to model combination is straight averaging, an approach often referred to as "pooling". It is easily shown that overall, pooling leads to more skilful and reliable forecasts than is possible with a single model. More elaborate techniques can involve recalibration of the probability distributions from the individual models, performance-weighting the models, or some combination of both. This talk outlines the evolution of IRI's approach to probabilistic seasonal forecasting from methods that we've tried, what we're using now, and finally the combination of techniques currently being developed as we work towards putting out a more flexible forecast product. Some of the techniques illustrated and evaluated in this talk will be categorical contingency table correction of probabilities, canonical variate ensembling, probabilistic regression methods, Bayesian ensembling, and analytical recalibration of both local and non-local model response. Pooling will serve as the methodological baseline. The relative enhancements in reliability and/or sharpness from the various approaches will be compared. Not all approaches benefit both the sharpness and reliability aspects of forecast quality.
H32C-02 INVITED
MERGING MULTIPLE CLIMATE MODEL FORECASTS FOR SEASONAL HYDROLOGIC PREDICTIONS
Skillful seasonal hydrologic predictions are required in water resource management, preparation for drought and its impacts, energy planning, and many other related sectors. In this study, a seasonal hydrologic ensemble prediction system is developed and evaluated over the Eastern U.S., with focus on the Ohio River basin. The seasonal hydrologic prediction system utilizes a hydrologic model (in this case the Variable Infiltration Capacity model) as the central element for producing ensemble hydrologic predictions of soil moisture, snow and streamflow with lead times up to 6 months. The uniqueness of this forecast system is in the method for generating ensemble atmospheric forcings for the forecast period. It merges seasonal climate predictions from multiple climate models with observed climatology in a Bayesian framework such that the uncertainties related to the atmospheric forcings can be reduced and better quantified. This framework also downscales the climate model forecasts to scales appropriate for hydrologic prediction and uses a rank structure of selected historical forcings to ensure that generated ensembles of daily meteorological forcings have reasonable patterns in space and time. Three types of forecasts were performed in the study: those using information from NCEP's Climate Forecast System (CFS), those using information from CFS and the European Union funded multi-model prediction project called DEMETER, and those based the Extended Streamflow Prediction (ESP) approach. Forecasts (CFS, CFS+DEMETER and ESP) were made with the system for the summer periods (May to October) for 1981 - 1999, and represent forecast information from one climate model, eight climate models and none, respectively. The differences in forecast skills between CFS, CFS+DEMETER and ESP reflect the improvement with the new forecast method against the current hydrological operational approach, which is based on ESP. The forecast for the summer 1988 shows very promising skill in precipitation, soil moisture and streamflow forecast over the Ohio river basin, especially with the CFS+DEMETER forecast. The evaluation over all 19 summer forecasts shows significant skill improvement with the new multi-model method during the first two months of the forecasts. The improvement is marginal to moderate when only CFS forecast is used. This study validates the approach of using seasonal climate predictions from dynamic climate models in hydrological predictions. It also shows the need for international collaborations to develop multi-model seasonal predictions. http://hydrology.princeton.edu/forecast
H32C-03
Predictability of U.S. Winter Precipitation: Role of ENSO state in Developing Multimodel Combinations
Recent research shows that operational climate forecasts obtained by combining different General Circulation Models (GCMs) have improved predictability in comparison to the predictability that could be obtained from a single GCM. In this study, we evaluate the skill of three GCMs in predicting the U.S. winter (December-February) precipitation conditioned on the state of El Nino-Southern Oscillation (ENSO). Using Nino3.4 as the conditioning variable, we show that the skill of GCMs in predicting the U.S winter precipitation is significant only when ENSO conditions exist. Under neutral ENSO conditions, predictability of three GCMs is statistically insignificant. Hence, we propose an algorithm for combining precipitation from multiple GCMs that considers the state of ENSO in developing multimodel ensembles of winter precipitation over the U.S. The approach basically identifies similar conditions or analogue years from the current state of Nino3.4 and then evaluates the average skill of the candidate GCMs during those conditions by computing the average Rank Probability Score (RPS). Multimodel ensembles of precipitation are then developed by drawing ensembles from each model in such a way that the model with low average RPS constitutes higher number of ensembles in the multimodel ensembles. The performance of multimodel ensembles is compared with individual model ensembles in predicting winter precipitation using various performance measures such as Rank Probability Skill Score (RPSS) and reliability plots.
H32C-04
Understanding uncertainties in hydrological models: Insights gained from a large ensemble of model structures
Multi-model ensembles are only valuable if there is independent information in the different models used to construct the ensemble. The critical question therefore is what is the nature of differences between models? To address this question we mix-and-match different architectures and different process parameterizations from different models to construct 79 hydrological models, all with different structure. These models were used to simulate streamflow in two basins in the USA: the Guadalupe River (Texas) and the French Broad River (South Carolina). The main insights gained from this exercise are [1] the differences in skill between models were larger in the Guadalupe River (drier basin) than the French Broad (wetter basin); [2] differences in model skill can be attributed to the choice of model structure -- the models with highest skill in the Guadalupe River were those that had low frequency variability in saturated areas (this was easiest to achieve when saturated area is controlled by lower zone storage); and [3] many models had similar errors for the same storm, suggesting the independent information in multiple models may be quite limited (or alternatively, there are large errors in model inputs). Further application of these models in different river basins will elucidate the independence between models, and determine which model structures are most suitable in specific environments.
H32C-05
Application of a Multi-Scheme Ensemble Prediction System and an Ensemble Classification Method to Streamflow Forecasting
Europe has experienced a number of unusually long-lasting and intense rainfall events in the last decade, resulting in severe floods in most European countries. Ensemble forecasts emerged as a valuable resource to provide decision makers in case of emergency with adequate information to protect downstream areas. However, forecasts should not only provide a best guess of the state of the stream network, but also an estimate of the range of possible outcomes. Ensemble forecast techniques are a suitable tool to obtain the required information. Furthermore a wide range of uncertainty that may impact hydrological forecasts can be accounted for using an ensemble of forecasts. The forecasting system used in this study is based on a multi-scheme ensemble prediction method and forecasts the meteorological uncertainty on synoptic scales as well as the resulting forecast error in weather derived products. Statistical methods are used to directly transform raw weather output to derived products and thereby utilize the statistical capabilities of each ensemble forecast. The forecasting system MS-EPS (Multi-Scheme Ensemble Prediction System) used in this study is a limited area ensemble prediction system using 75 different numerical weather prediction (NWP) model parameterisations. These individual ‘schemes' each differ in their formulation of the fast meteorological processes. The MS-EPS forecasts are used as input for a hydrological model (HBV) to generate an ensemble of streamflow forecasts. Determining the most probable forecast from an ensemble of forecasts requires suitable statistical tools. They must enable a forecaster to interpret the model output, to condense the information and to provide the desired product. For this purpose, a probabilistic multi-trend filter (pmt-filter) for statistical post-processing of the hydrological ensemble forecasts is used in this study. An application of the forecasting system is shown for a watershed located in the eastern part of Germany for the severe flooding of the river Elbe in August of 2002.
H32C-06
Accounting for Uncertainty Propagation: A Streamflow Forecasting Framework using Multiple Climate and Hydrological Models
Water resources planning and management efficacy is subject to capturing inherent uncertainties stemming from climatic and hydrological inputs and models. Accounting for and properly dealing with these propagating uncertainties remains a formidable challenge. Streamflow forecasts, critical in reservoir operation and water allocation decision-making, fundamentally contain uncertainties arising from assumed initial conditions, model structure, and modeled processes. 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 (GCM), multiple regional climate models, and numerous hydrological models to improve streamflow forecasting and characterize system uncertainty 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 GCM and regional models, including dynamical and statistical models, are integrated with the Sacramento Soil Moisture Accounting and SMAP (Soil Moisture Accounting Procedure) hydrological models. Precipitation hindcasts from the GCM are downscaled via the regional models and fed into the hydrological models, producing streamflow hindcasts. Multi-model ensemble combination techniques include pooling, least squares regression, and a kernel density estimator to evaluate streamflow hindcasts and assess structural uncertainty of climate and hydrological models; the latter technique exhibits slightly superior skill compared to any single coupled model ensemble hindcast.
H32C-07
Ensemble Forecasts for Water Management Optimization Procedures (EOP)
Science and technology support water management through weather forecasts, hydrologic forecasts, and water management models. In modern water management practice these three tools and their data are integrated into seamless "Decision Support Systems" with convenient user interfaces. This paper discusses the last step in such systems for reservoir operations management. This is an optimization step that incorporates uncertainties into suggestions for making the best possible decisions. Ensemble weather forecasts typically provide a set of alternative equally likely weather sequences over the forecast period. These are used as inputs to hydrologic models to provide a suite of hydrologic time series that represent equally likely alternative water supply and flood forecasts that are conditional on current weather and watershed initial conditions. The decision making process at the water management level includes the following steps: evaluate risks, develop contingency plans, implement mitigative actions, determine optimal operating decisions. These steps each require a probabilistic analysis that may be optimized with an appropriate procedure for the specific application. If the right programs and data management tools are available, and readily accessible, hydrologic ensemble forecasts can provide inputs to an Ensemble Optimization Procedure (EOP) to support decision making. This paper discusses practical implementation issues and demonstrates three EOP methods for scheduling operation of a hypothetical hydroelectric reservoir.