Second International Conference on Subseasonal to Seasonal Prediction (S2S) and
Second International Conference on Seasonal to Decadal Prediction (S2D)

Daily news

Day 2: A look at modeling and forecast challenges and advancements

For the second day of meetings, participants learned about modeling issues, ensemble predictions, and forecast information available from the subseasonal to decadal timescales. The concurrent sessions continued for both the S2S and S2D topics; below are some highlights.

Subseasonal to Seasonal Conference Highlights

Operational forecast centers are striving to increase the accuracy of forecasts. And while each year progress is made, challenges remain in the methods and initialization of the models, increasing or identifying the appropriate resolution, determining the ensemble size, and simulating and understanding uncertainties. The use of large ensembles show promise in identifying some of these challenges. Many talks also described how S2S predictions can be used for extreme and high-impact events.

The first session of the day was on modeling issues, chaired by Anca Brookshaw, ECMWF. Yuhei Takaya, Japan Meteorological Agency, presented a keynote lecture on the art and science and goes into designing operational forecast systems. He reviewed current practices and approaches, including the pros and cons, at operational centers. He noted that larger ensembles are better, but there is diminishing returns to skill improvement. He also showed ocean and atmosphere initialization shocks and model drifts that appear on short timescales – often from using different models, versions, or configurations – and that a possible solution is to use coupled data assimilation. Another major challenge for coupled S2S modeling is the lack of observations, such as sea ice and land.

As part of the ensemble predictions and forecast information session, Laura Ferranti, ECMWF, was interested in knowing how far in advance we can predict changes in large-scale flow that influences severe cold weather over Europe and other extreme events. Looking at the predictability of the North Atlantic Oscillation (NAO) and blocking, she found that S2S systems exhibit useful skill well beyond 10 days, which provide the potential for early warning systems. She also emphasized the need to collaborate with users to get a clear idea on what they want to know, and only with that can we make a useful product for longer range forecasts.

The day was wrapped up with a group discussion on the S2S Prediction Project database, with the leads looking for users’ experience in accessing the data and ideas for improvement. Some of the discussion centered on accessibility. Specifically, the S2S Prediction Project releases their forecasts on a three-week delay (due to embargoed data), but with the emergence of other projects, like SubX, that have immediate, open availability of model outputs that are useful when developing forecasts products, the three-week delay may need to be revised and renegotiated.

Thomas M. Hamill, Cécile Penland, both from NOAA ESRL, and Cristiana Stan, George Mason University, also chaired sessions throughout the day.

Additional highlights include: 

• Many operational centers are subjected to practical constraints dictated by operational activities (timeliness, costs, priorities), and they face a challenge of identifying suitable compromises, which have no detrimental effect on the skill of predictions. (multiple speakers)
• Multiple operational centers are moving towards a unified, coupled forecast system that can work across timescales from days (or shorter) to seasons (or longer). As forecast systems become more complex and more expensive computationally, while evidence in favor of the need for large ensembles of simulations mounts up, ideas on fundamental changes to the modeling paradigm are being tested in research mode (e.g., single-precision modelling). (multiple speakers)

• There is an ongoing debate about whether high-resolution models are necessary for S2S predictions. Some argue that inexpensive options – linear inverse models (for seasonal ranges) and low-resolution models (25–50km) with stochastic parameterisations (for subseasonal ranges) – are competitive candidates to high-resolution global climate models. However, it was acknowledged that the conclusion may be different when the emphasis shifts to medium range (~10 days) forecasts, where deterministic aspects are of high interest. (Prashant Sardeshmukh, University of Colorado/CIRES & NOAA Earth System Research Lab)   
• A common topic throughout the day was the Madden-Julian Oscillation (MJO), which depends crucially on moisture budget and geographical distribution of moisture. Thus, dry biases in models adversely affects MJO representation and prediction, especially when it comes to propagation through the maritime continent and into the Pacific warm pool. (multiple speakers)
• Using a verification framework for evaluating subseasonal South American precipitation prediction, which assesses weekly hindcast/real-time forecast quality, researchers were able to address the quality of prediction attributes (i.e., association, accuracy, reliability, and resolution). (Caio Coelho, Weather Prevision Center and Climate Studies/INPE, Brazil)  
• Researchers have found that the MJO and Quasi-Biennial Oscillation (QBO) can be used to improve the subseasonal predictability of atmospheric rivers, in addition to skillfully forecasting other anomalous extreme weather events, by using an index to quantify the diagonal “stripey-ness” of composite plots of anomalies. (Elizabeth Barnes, Colorado State University)
• Improvement in model skill is possible through different techniques, such as using a statistical model based on lagged composite fields associated to teleconnection patterns or a multimodel weighted calibrated ensemble. However, the improvement is only slight. (multiple speakers)
• The ability of statistical or dynamical models to skillfully predict surface air temperature on S2S timescales is regionally dependent – such as East Asia, which sees the highest impact from teleconnection patterns which are upstream. Additionally, the skill of models seems to be better when forecasting minimum temperature and minimum-temperature-related heat waves than maximum temperature or heat waves. (Changhyun Yoo, Ewha Womans University, Korea; Lauriane Batté, National Centre for Meteorological Research, France)
• Conditional states of the climate system offer opportunities for skillful forecasts over limited spatio-temporal subsets of the observed record. (Violeta Toma, Climate Forecast Applications Network)

Seasonal to Decadal Conference Highlights

On the second day of the conferences, participants in the S2D track reassembled to continue with presentations on the theme of “Mechanisms of S2D Predictability.” Speakers discussed important sources of S2D hydroclimate predictability, including remote ocean forcing (such as ENSO, Pacific Decadal Oscillation (PDO)/Interdecadal (IPO), tropical Atlantic SST, Atlantic meridional overturning circulation (AMOC)/Atlantic multidecadal variability (AMV), tropical monsoon convection (such as the Indian summer monsoon) exerting atmospheric teleconnection, and local land memories mainly due to soil moisture and vegetation. The speakers noted that realistic initializations and simulations of land-atmosphere interactions and processes, such as soil moisture and vegetation sensitivity, can provide improvement in S2D prediction skill. The influence of ENSO on climate variability and S2D prediction skill in many parts of the globe was seen as highlighting the importance of improving ENSO prediction. It was noted that recent advances in global seasonal forecast models tend to breach the ENSO springtime predictability barrier.

The second session was on “Modeling Issues in S2D Prediction.” The opening keynote by Wolfgang Müller (Germany's National Meteorological Service, Deutscher Wetterdienst) presented on the operationalization of seasonal-to-decadal climate predictions with the Max Planck Institute Earth System Model (ESM). Müller underlined that preparing an ESM for operational S2D predictions requires joint effort of different communities, including S2D predictions, climate modeling, and data assimilation. He addressed prominently the problem of model biases, a topic thoroughly discussed throughout the session. Given that model biases are not expected to vanish any time soon, speakers suggested different ways to reduce or overcome model biases, for example by exploring innovation on and understanding of model bias, process-oriented model diagnostics and biases in variability, the evolution of model biases with lead time, and by comparing the evolution to historical/unconstrained model simulations and/or Atmospheric Model Intercomparison Project (AMIP)-experiments.

Tuesday afternoon speakers highlighted the improvement of skill in their climate prediction models, but also noted that newer model versions with different configurations can lead to decreased skill in some areas. Further techniques to improve predictive skill were introduced. For example, it was reported that initialization with reanalyses assimilating upper atmosphere observations can improve wintertime NAO skill of seasonal forecasts – a topic strongly related to correct QBO initial conditions. The introduction of an ensemble dispersion filter to the MiKlip – a German research project on decadal predictions – improved the forecast skill for years 2–5. Ocean filtering and atmospheric variations, such as the MJO, were presented both as important contributors to initialization. However, filtering is not recommended for the initialization of the MJO.

Tuesday’s last three talks opened the theme of “S2D Ensemble Predictions and Forecast Information.” A keynote by Stephen Yeager, NCAR, addressed near-term hydroclimate outlooks, based on the Community Earth System Model (CESM) Decadal Prediction Large Ensemble. Results from this large ensemble showed that decadal land-based precipitation skill is complex, varying with region, season, lead-time, ensemble size, inclusion or non-inclusion of trend, spatial filtering, and other factors. Promising results were, however, presented for the Sahel region, Northern Europe, and Eurasia, specifically. Additionally, SST skill improves with lead-time, resulting in improved land-based precipitation skill. Further presentations addressed multi-year prediction of La Niña events, where exact results depend on underlying model dynamics, but were shown to hold skill out to 24 months ahead. An open question remains whether these results can transfer to warm events.