Ensemble Forecasts for Weather and Seasonal Climate II
Presiding: T Krishnamurti, Florida State University; B Rajagopalan, University of Colorado
A14A-01 INVITED 15:30h
Multi-Model Ensembling: A Comparison of Approaches
A global climate forecast system should be as accurate as possible with probabilities that are as reliable as possible. Individual general circulation models (GCMs) can differ substantially in their seasonal forecast skill due to differences in parameterizations and even in the structure of the models' dynamical cores. A few regions of high predictability exist, for which most, if not all, models simulate the interannual variability accurately when forced with appropriate surface conditions. For most of the continental areas, however, the variability is not so straightforward or deterministic, and in these cases the details of the GCM formulation greatly influences the model's ability to simulate the characteristics of the variability. No one model is best. Thus the value of combining the predictions from several models is clear. Currently a number of approaches exist for combining probabilistic predictions from GCMs. Many consider some measure of model skill in determining the regional weighting given to a particular GCM. In this presentation, several approaches are compared and contrasted; each constitutes a somewhat different philosophy in model combination. Two of the approaches, straight pooling and prediction convergence, do not consider historical model performance. Two other approaches, Bayesian and canonical variate, do incorporate information regarding past performance of the individual models. All multi-model ensembling (MME) approaches yield an overall improvement on skill relative to the performance of the individual models. Occasionally, the MME even outperforms the best model over a particular region. Although the different MME approaches give seasonal forecasts that are more similar to each other than were the constituent model predictions, considerable differences still remain. Differences in skill are evident in the different MME approaches, with the performance-based approaches generally scoring higher. Differences are found also in the sharpness of the forecast probabilities and the amount of area over which non-climatological probabilities are indicated.
A14A-02 15:45h
Skill of Forecast Methodologies for US Seasonal Climate
This study appraises prospects for US seasonal forecast skill by diagnosing various dynamical modeling systems. Parallel evaluation of an empirical forecast model provides a performance baseline. Results are drawn from hindcast applications of each system for the 20-year reference period 1981-1999. Two classes of dynamical model systems are studied. One consists of atmospheric models subjected to specified sea surface temperatures comprising the so-called two-tier prediction system. Another consists of coupled atmosphere-ocean models comprising the so-called one-tier system. Multi-models are available for each methodology, and ensemble experiments are generated for every model. We discriminate between the performances in different verions of each system. Two generations of the two-tiered dynamical systems are available: a multi-model suite of four GCMs developed in the 1990s and a comparable suite of their subsequent model versions released in recent years. For the one-tier, a seven-model suite of coupled hindcasts is derived from the DEMETER project, and their performance is compared to a recent coupled forecast system developed in support of U.S. seasonal climate forecasting. The presentation focuses on two questions: Is U.S seasonal forecast skill advancing with newer generation modeling systems? and Is the potential predictability of U.S. seasonal forecasts being fully realized? The skill and predictability addressed in our study is that resulting principally from air-sea interactions and the accompanying atmospheric teleconnections linking U.S. climate with sea surface temperatures. We focus on deterministic measures of that skill, using a temporal anomaly correlation metric for the 20-yr verification period.
A14A-03 16:00h
A Performance Comparison of Coupled and Uncoupled Versions of the Met Office Seasonal Prediction GCM
The benefits of using coupled GCMs over two-tier AGCM systems for seasonal prediction to 6-month range will be explored using 43-year, 9-ensemble hindcast datasets generated as part of the EU project DEMETER. The CGCM and AGCM systems employed share the same atmospheric component and a performance comparison therefore provides insight into the skill benefits available from coupling atmosphere and ocean models. The two-tier AGCM is forced with predicted Sea Surface Temperature (SST) based on persistence of observed SST anomalies (SSTA). Analysis will be focussed on global and regional comparisons of long-term skill for probabilistic prediction of 2m temperature. In the tropics CGCM benefits are seen, as expected, in the tropical east Pacific and are associated with superior CGCM skill (relative to persisted SSTA) in predicting ENSO SST anomalies. However substantial benefits are also seen throughout the tropical belt in seasons associated with the peak and decay of ENSO activity, such benefits arise from representation of lagged teleconnection responses to ENSO in the tropical Atlantic and Indian Oceans. CGCM benefits are also found in extratropical regions. In this regard an investigation of CGCM skill benefits for prediction of spring season temperature in the European region will be described in which improved skill appears to derive from coupled model representation of linkage between the well documented North Atlantic SSTA tripole pattern and the North Atlantic Oscillation (NAO). This result provides encouraging evidence that use of coupled GCMs offers prospects for improving seasonal prediction in the extratropics through representation of coupled ocean-atmosphere processes in extratropical ocean basins, as well as through indirect impacts from improved prediction of ENSO and associated teleconnections.
A14A-04 16:15h
Ensemble forecasts and seasonal precipitation tercile probabilities
Seasonal forecasts of precipitation and near-surface temperature are often issued in terms of expected probabilities of tercile categories. A simple method of estimating precipitation tercile probabilities from seasonal ensemble forecasts is to use relative ensemble frequencies, that is, to count the number of ensemble members in each category. Such a nonparametric estimate is affected by sampling error, especially for modest sized ensembles like those used in seasonal forecasting. Previous work has shown the benefit of estimating tercile probabilities from parametric distributions fit to ensemble output, though variables like precipitation with nonnormal distributions provide challenges. Here we compare parametric and nonparametric methods of estimating tercile probabilities to an empirical approach using a generalized linear model. The empirical approach allows us to quantify the ensemble information necessary to determine tercile probabilities, and in particular, to determine the relative importance of ensemble mean and spread on tercile probabilities. Results are obtained using idealized models and output from general circulation atmospheric models.
A14A-05 16:30h
Multi-model Ensemble Forecast of Spring Seasonal Flows in the Gunnison River Basin
Water managers need ensemble streamflow forecasts so as to obtain ensemble of decision variables for optimal management and planning of water resources. This is more so the case in the western US. Large scale ocean-atmospheric features in the Pacific, in particular, have a significant impact on the variability of hydroclimatology in the western US. As a result, any ensemble forecast method should include this large-scale forcing information. Furthermore, the streamflows are required at several locations in a basin and the ensemble forecast method should preserve the spatial variability as well. To this end, we propose a multi-model ensemble forecast approach that has the following four steps: (i) Principal Component Analysis is performed on the spatial streamflows to identify the dominant modes of variability (ii) Large scale ocean-atmospheric predictors are identified for the dominant modes, (iii) Objective criteria, Generalized Cross Validation (GCV) is used to select a suite of candidate models, and (iv) Ensembles of forecast of the dominant modes and consequently the spatial flows are issued from the candidate models. The forecast model is local polynomial based nonparametric regression approach. The main advantage of this approach is that by forecasting ensembles for a suite of candidate models, model uncertainty is better captured. In addition, forecasting the dominant modes and then obtaining the flows helps in preserving the spatial variability. The utility of the framework is demonstrated in forecasting Spring streamflows at six locations in the Gunnison river basin. The approach exhibited substantial cross-validated skill.
A14A-06 16:45h
Performance of the Florida State University Hurricane Superensemble During 2004
The performance of the FSU hurricane superensemble (SE) is evaluated for the active 2004 Atlantic-basin season. During this memorable season in which five hurricanes made landfall over the US -- including two major hurricanes striking Florida -- the FSU multi-model SE established itself as one of the very best forecasting tools in use at the National Hurricane Center (NHC). This statistical post-processing technique utilizes the NHC suite of real-time model predictions of hurricane track locations (lat/lon pairs) and intensity values. Unequal weights are then applied to each forecast model for each 12-hour forecast interval. These weights are pre-determined through a training phase, during which a statistical relation is obtained among past model forecasts and observed data using a multiple linear regression approach. This year, a new non-linear technique was implemented for the first three days of the track forecast. This had a positive impact in improving the SE performance. For the 2004 season as a whole, the track SE had comparable RMS errors to the GUNA consensus model through forecast hour 108, meanwhile outperforming all other model forecasts, including the ensemble mean. The strength of the track SE was evident in the longer range, where on average it showed a 5.5% improvement over the best model at 120 hours (122 cases). When examining the cases where the intensity of a cyclone was at least minimal hurricane strength, the SE demonstrated an 11.1% improvement over the best model at 120 hours (92 cases). Additionally, the average 72-hour SE intensity forecast exhibited a 17.4% improvement over the best intensity model (200 cases). Moreover, the superensemble performed very impressively with the long-track Cape Verde-type hurricanes. The landfall predictions of these hurricanes were quite accurate even well before 72 hours. For instance, the SE predicted well the landfall point of Frances over the Florida east coast about 84 hours in advance. With Ivan, the FSU SE did not forecast any landfall over peninsular Florida while the member models incorrectly did so. Accordingly, the SE held a clear advantage at the 120-hour lead-time with average errors for Ivan about 170 km less than the best model (35 cases). Also, the looping of Jeanne was well-predicted even when the member model tracks had a large spread. Upon examining the forecasts of a particular tropical cyclone over its entire life cycle, it is evident that the spread of the SE is much less than the other models. This indicates the greater forecast consistency and reliability of the FSU SE.