A New Way to Forecast

ECMWF’s core activity is to provide useful weather forecasts for the world using a global numerical model. Whilst this goal has been successfully met, there continue to be limitations. Notably when forecasting for points the global model output suffers from representivity errors, and there are also biases. By representivity errors we mean sub-grid variability: this might be due to (e.g.) “randomness” in rainfall on a convective day, or due to the influence of unrepresented surface features (e.g. topography) on weather at points. By biases we generally mean the average difference, over many cases, between short range gridbox forecasts, intrinsically provided by the model, and the average of evenly-spaced high-density observations within the gridbox.

So to meet the key requirement of most customers, which is providing point forecasts, there is a need for adjustment. The first ‘zeroth order’ approach to this is interpolation (as on the ECMWF Meteogram product), the second is dynamical downscaling via a high resolution limited area model fed by global model boundary conditions, which is very costly and which is not performed at ECMWF, and the third is statistical downscaling. Statistical downscaling has generally been applied to sites where observational data exists, and as such needs long training periods using the same model version.

Here a new approach to statistical downscaling will be described. This uses observations but is not site specific, so can fill data voids. Put simply, an observation in (e.g.) central Asia might help to improve the forecast for a site in (e.g.) the central US. A large observational dataset helps, so we use global SYNOPs. High density data is not necessary, nor is a long training period. The method relies crucially on physical understanding of factors affecting point weather. For example light steering winds on convective days favour enhanced sub-grid variability in rainfall totals. Factor values are extracted from model output (and from geographical datasets) to define gridbox scenarios, and these scenarios are used as the basis for post-processing a gridbox forecast into a pointwise probability density function (pdf) of possible outcomes, for that scenario, in that gridbox. Pdfs address both mean bias and sub-grid variability. The final forecast is naturally probabilistic. It is derived by first post-processing each ensemble member independently to give point pdfs, and by then summing those pdfs.

At ECMWF the above strategy has been used in the development of a global flash flood prediction system, based on rainfall post-processing. Screen temperature downscaling has also been explored. Together these can pave the way for introduction of a post-processed Meteogram product, that should be more reliable and less biased. Meanwhile the research stage has been yielding fresh insights into model behaviour. There is also potential to apply these very cost-effective methods to monthly, seasonal and climate change forecasts. In tandem with the 2017 AMS meeting theme, and with the desire of any good synoptician, observational data here is being used to its utmost.