Seminar: Dr. S. Lovejoy and Dr. L. Del Rio Amador
Please join us as we welcome Dr. S. Lovejoy and Dr.聽 L. Del Rio Amador from the department of Physics at 苹果淫院 for their seminar titled "Harnessing butterflies for improved monthly, seasonal, and interannual forecasts". Coffee will be served.
Abstract
Over the past ten years, a key advance in our understanding of atmospheric variability is the discovery that between the weather and climate regime lies an intermediate 鈥渕acroweather鈥 regime, spanning the range of scales from 鈮 10 days to 鈮30 years (in the anthropocene; it is longer in the pre-anthropocene).聽聽 Macroweather statistics are characterized by two fundamental symmetries: scaling and the factorization of the joint space-time statistics.聽 In the time domain, the scaling has low intermittency with the additional property that successive fluctuations tend to cancel.聽 In space, on the contrary the scaling has high intermittency corresponding to the existence of different climate zones.聽聽
These properties are fundamental for macroweather forecasting.聽 For example:
- The temporal scaling implies that the system has a long range - indeed elephantine - memory that can be exploited for forecasting.
- The low temporal intermittency implies that mathematically well-established (Gaussian) forecasting techniques can be used.
- The statistical factorization property implies that although spatial correlations are large, that they are not useful in making forecasts.
These properties can be directly exploited by the Stochastic Seasonal and Interannual Prediction System (StocSIPS).聽 StocSIPS is a straightforward, highly efficient forecasting system that makes global, monthly, seasonal and interannual forecasts.聽 Using hindcasts, we compare StocSIPS with Environment Canada鈥檚 CanSIPS model, finding that most of the earth, for horizons beyond about one month, that StocSIPS is significantly more accurate.聽
StocSIPS鈥 advantages include:
- Convergence to the real 鈥 not model - climate: The key to StocSIPS skill is the forecasting module that uses past data 鈥 and the huge memory in the system - to ensure that the forecast converges to the real world climate.
- Speed: In order to get good statistics, conventional seasonal to annual forecasts typically re-forecast over ten to twenty realizations, each time using slightly different initial data often taking the equivalent of hundreds of thousands of CPU hours on the world鈥檚 fastest computers.聽 In comparison, StocSIPS uses only a few minutes of CPU time to directly calculate the statistics of an infinite number of realizations.
- No data assimilation: StocSIPS can directly forecast either gridded or individual station data, there is no need to transform the input data to make it digestible by the numerical model; StocSIPS avoids complex data 鈥渁ssimilation鈥 techniques.聽聽
- No ad hoc post processing: The raw temperatures and precipitation rates forecast by conventional models have unrealistic variability.聽 This is usually 鈥渃orrected鈥 using complex ad hoc post processing algorithms that use hindcasts to incorporate past information in order to make the forecasts more realistic.聽 StocSIPS uses only past information with a theoretically justified forecast procedure.
- No need for downscaling: Conventional models have pixels of 100,000 km2 or more in size and must be 鈥渄ownscaled鈥 to adapt them to local conditions.聽聽 Whenever long station temperature series are available, StocSIPS can forecast them directly.