An evaluation of WRF model output statistics techniques in eastern Tennessee

Models of surface winds are key tools for predicting the atmospheric transport and dispersion of toxic and other hazardous materials. The National Oceanic and Atmospheric Administration (NOAA) Air Resources Laboratory/Atmospheric Turbulence and Diffusion Division (ARL/ATDD) runs a Weather Research and Forecasting (WRF) mesoscale model over an eastern Tennessee domain for dispersion prediction purposes. The WRF model can resolve larger terrain features in the domain, such as the Cumberland Plateau and the Great Smoky Mountains, but it cannot resolve the smaller ridges in the Tennessee River Valley. These small-scale features cause biases in WRF forecasts for that area. Model Output Statistics (MOS) based on the Regional Air Monitoring and Analytical Network (RAMAN) were developed to correct for those biases. Two MOS techniques were evaluated: the traditional screening regression and the M5 Model Tree. The M5 uses a decision tree to partition the full domain of predictor variables into smaller subdomains over each of which it fits a linear regression to the data within. Being piecewise linear, M5 promises a better fit. MOS models were compiled for 16 sites, 2 seasons, and 4 diurnal time periods. Six parameters were predicted: temperature, wind speed, wind components (u,v), standard deviation of wind direction, and probability of calm. The two statistical models' prediction errors were estimated by cross validation, allowing comparison with each other and with the direct WRF output. Both bias error and mean absolute error were compared using bootstrapped confidence intervals to determine significance. The M5 Model Tree was found to provide significantly reduced error in its predictions compared to the screening regression, and both MOS models significantly outperformed the WRF output. For wind speed, the WRF had a significant bias and large variance while the MOS models corrected for the bias and reduced the variance as well. The WRF did not have a bias in wind direction, but it did have a larger variance, which was also reduced in the MOS. The MOS models for probability of calm were found to be unbiased and skilled, the M5 Model Tree having the higher skill. The MOS is now running operationally with live verification of the predictions. The verification has so far shown a diurnal trend in temperature errors, low error in wind speed predictions, and a mix of large and small errors in wind direction.