An optimal model output calibration algorithm suitable for objective temperature forecasting

An optimal model output calibration (MOC) algorithm suitable for surface air temperature forecasts is proposed and tested with the National Centers for Environmental Prediction Regional Spectral Model (RSM). Differing from existing methodologies and the traditional model output statistics (MOS) technique, the MOC algorithm uses forecasts and observations of the most recent 2-4 weeks to objectively estimate and adjust the current model forecast errors and make refined predictions. The MOC equation, a multivariate linear regression equation with forecast error being the predictand, objectively screens as many as 30 candidates of predictors and optimally selects no more than 6. The equation varies from day to day and from site to site. Since it does not rely on long-term statistics of stable model runs, the MOC minimizes the influence of changes in model physics and spatial resolution on the forecast refinement process. Forecast experiments were conducted for six major urban centers in the Tennessee Valley over the period of 27 June to 30 July 1997. Surface air temperature forecasts out to 72 h were produced based upon RSM runs initialized from 0000 UTC observations. Performance of the MOC for minimum and maximum temperature forecasts was assessed by determining mean forecast error (BIAS), mean absolute error (MAE), and root-mean-square errors (rmse) for both the MOC-adjusted and nonadjusted RSM output. The same statistical measures for Nested Grid Model-MOS forecasts over the experiment period were also provided for instruction. A skill score was calculated to demonstrate the improvement of refined forecasts with the MOC over the RSM. On average for the six sites, reduction of forecast errors by the MOC ranged from 58% to 98% in BIAS, 40% to 52% in MAE, and 33% to 46% in rmse. It also showed that the error frequencies of the refined forecasts had Gaussian distributions with the peak centered around zero. The error bands were narrower using the MOC and there were decreases in large forecast errors, especially during the first 48 h. The Wilcoxon signed-rank test was performed to verify that populations of the forecast errors before and after the MOC adjustment were statistically far enough apart to be distinct at a high significance level. Forecast experiments were also conducted to address the issue of sensitivity of the MOC by varying the length of the time series used in deriving the MOC equation. It was found that the mean biases of the refined forecasts s