Statistical downscaling of 2 m air temperature forecasts from the Eta Model at 40 km resolution were carried out for 7 sites located in Northeast Brazil. The downscaling method was the Model Output Calibration (MOC), which is based on the linear regression between the forecast error and model output variables. One feature of the scheme is the use of short training period, in this work only 28 days was the chosen period. The model output variables with correlation higher than 0.6 were selected as predictors of the regression equation. The scheme was tested for one-month hourly forecasts in September 2003. Model systematic errors were strongly reduced to values below |0.5| °C on average. The root mean square errors, which initially were of the order of 2 °C, were reduced to about 1 °C. The linear correlation between the observations and Eta forecasts were generally kept the same after applying the downscaling scheme. Temperature forecasts with greatest errors in the amplitude of the diurnal cycle were corrected accordingly.
Statistical downscaling; Eta Model; temperature forecasts
