Development of a methodology to reduce the near-surface prediction error in mesoscale atmospheric models.

摘要

Mesoscale meteorological models are used extensively to provide high-resolution near-surface meteorological fields for various scientific and operational applications (e.g., local-scale weather forecasting and regional-scale air quality studies). During the last few decades, there have been serious efforts to improve accuracy of model outputs using better initialization procedures, increasing grid resolution, assimilating both conventional and unconventional data (e.g., ground- and space-based remote sensing) and better physical parameterization schemes. In spite of these improvements, near-surface model forecast error still remains a difficult problem to correct, mainly due to inaccuracies in representation of the air-surface interactions.; Methods for further improvement of a near-surface prediction and for a better interpretation of model outputs have been investigated in this thesis. They include a more accurate representation of surface physics, a study of a model-specific sensitivity to surface properties initialization, and an application of a neural network based technique to reduce near-surface prediction error relying on analysis of prior model output and observations. Results of this study indicated that neural network technique can improve the near-surface forecast of a mesoscale model. Furthermore, for mesoscale model forecast error reduction, the neural network technique seems to be a promising method and therefore this technique should be investigated in more details, forming the basis for future research.