• English 
  • Spanish 

Statistical downscaling and local weather forecast

logo_downscaling.png Short description: Statistical and machine learning techniques applied to local weather forecast by adapting the prediction of numerical models using statistical relationships obtained from historical records

The demand for high-resolution weather forecasts (e.g. local daily prediction of precipitation in Madrid, Spain) is continuously increasing in a variety of socio-economic impact sectors, including turism, agriculture, energy, health, and insurance. However, the resolution of Global Circulation Models (GCMs) is currently constrained by computational and physical reasons to 200 km (for climate change predictions), 100 km (seasonal forecast), 50 km (medium-range 3-15 days), 5-10 km (short-range 1-3 days). In order to increase the spatial resolution of these predictions at the different lead times, a number of statistical downscaling techniques have been developed in the last decades.

Statistical downscaling techniques combine the information of retrospective GCM analysis/forecasts databases with simultaneous local historical observations to infer statistical relationships between the low-resolution GCM fields and the high-resolution observed records (usually surface variables such as precipitation or temperature). Nowadays there are a large number of different methods, which are usually classified in three categories: transfer functions (linear, neural networks, etc.), analogs and weather typing, and weather generators [see Benestad, R. et al. (2008) Empirical-Statistical Downscaling, World Scientific Publishers. pdf copy].

The skill of these methods varies depending on the variable, season and region, with the latter variation dominating [see e.g. Schmidli, J. et al., Statistical and dynamical downscaling of precipitation: An evaluation and comparison of scenarios for the European Alps, Journal of Geophysical Research, 122, D04105, 2007]. Thus, for each particular application and case study, an ensemble of statistical downscaling methods needs to be tested and validated to achieve the maximum skill and a proper representation of uncertainties.

Basic Reading:

  • R.L. Wilby and T.M.L. Wigley (1997) Downscaling general circulation model output: a review of methods and limitations. Progress in Physical Geography, Vol. 21, No. 4, 530-548 link
  • R. Huth (1999) Statistical downscaling in central Europe: Evaluation of methods and potential predictors, Climate Research, 13, 91-101 pdf

It is crucial to follow a minimum number of recommendations when using this methodology to downscale GCM outputs for different lead times:

  • Climate change scenarios: IPCC recommendations for the application of statistical downscaling techniques in climate change scenarios. Good practice guidelines from the ENSEMBLES portal for probabilistic regional climate information for impacts' assessments.
  • Seasonal forecast: ENSEMBLES deliverable about recommendations for the application of statistical downscaling methods to seasonal-to-decadal hindcasts.

Activities of the Santander Meteorology Group:

  • New statistical downscaling techniques based on data mining methods: bayesian networks, neural networks, etc., for different time scales: short-range, medium-range, seasonal forecast, climate change scenarios.
  • Adaptation of statistical downscaling methods for Ensemble Forecasts Systems.
  • Multi-site downscaling.
  • Regionalization of climate change scenarios.
  • Web tools for statistical downscaling.

The Statistical Downscaling (SD) Portal:

Statistical downscaling is nowadays a mature and complex multi-disciplinary field involving a cascade of different scientific tools to access and process large amounts of heterogeneous data. Therefore, interactive user-friendly tools are necessary in order to ease the downscaling process for end users, thus maximizing the exploitation of the available predictions. The SD Portal has been designed following an end-to-end approach in order to transparently connect data providers and end users. To this aim, Internet and distributed computing technologies have been combined together with statistical tools to directly downscale GCM outputs to the regional or local scale required by impact applications. Thus, users can test and validate online different methods (regression, neural networks, analogs, weather typing, etc.) working on a Web browser, not worrying about the details of the techniques used or the data accessed. The portal is part of the ENSEMBLES EU-funded project.
-> http://www.meteo.unican.es/ensembles

People: , , , , , , , ,