Articles | Volume 14, issue 3
Nonlin. Processes Geophys., 14, 211–222, 2007
https://doi.org/10.5194/npg-14-211-2007
Nonlin. Processes Geophys., 14, 211–222, 2007
https://doi.org/10.5194/npg-14-211-2007

  25 May 2007

25 May 2007

Prediction of minimum temperatures in an alpine region by linear and non-linear post-processing of meteorological models

E. Eccel1, L. Ghielmi1, P. Granitto2, R. Barbiero3, F. Grazzini4, and D. Cesari4 E. Eccel et al.
  • 1IASMA Research Centre – Natural Resources Department, Via E. Mach, 1 – 38010 San Michele all'Adige (TN), Italy
  • 2Instituto de Física Rosario Conicet UNR Bv. 27 de Febrero 210 bis 2000 Rosario, Argentina
  • 3Autonomous Province of Trento – Meteotrentino, Department of Civil Protection, Via Vannetti, 41 – 38100 Trento, Italy
  • 4ARPA-SIM Emilia-Romagna, Viale Silvani 6, 40122 Bologna, Italy

Abstract. Model Output Statistics (MOS) refers to a method of post-processing the direct outputs of numerical weather prediction (NWP) models in order to reduce the biases introduced by a coarse horizontal resolution. This technique is especially useful in orographically complex regions, where large differences can be found between the NWP elevation model and the true orography. This study carries out a comparison of linear and non-linear MOS methods, aimed at the prediction of minimum temperatures in a fruit-growing region of the Italian Alps, based on the output of two different NWPs (ECMWF T511–L60 and LAMI-3). Temperature, of course, is a particularly important NWP output; among other roles it drives the local frost forecast, which is of great interest to agriculture. The mechanisms of cold air drainage, a distinctive aspect of mountain environments, are often unsatisfactorily captured by global circulation models. The simplest post-processing technique applied in this work was a correction for the mean bias, assessed at individual model grid points. We also implemented a multivariate linear regression on the output at the grid points surrounding the target area, and two non-linear models based on machine learning techniques: Neural Networks and Random Forest. We compare the performance of all these techniques on four different NWP data sets. Downscaling the temperatures clearly improved the temperature forecasts with respect to the raw NWP output, and also with respect to the basic mean bias correction. Multivariate methods generally yielded better results, but the advantage of using non-linear algorithms was small if not negligible. RF, the best performing method, was implemented on ECMWF prognostic output at 06:00 UTC over the 9 grid points surrounding the target area. Mean absolute errors in the prediction of 2 m temperature at 06:00 UTC were approximately 1.2°C, close to the natural variability inside the area itself.

Download