Predicting the Momentum Flux-Profile Relationship from Macro Weather Parameters

Sammanfattning: The study of climate heavily relies on simulations. For efficiency reasons, many important phenomena cannot be simulated, and have to be parametrized, i.e. their effect must be described based on macro parameters. The turbulence resulting from the interaction between the atmosphere and the surface is an example of such phenomena. One of the quantities of interest arising from turbulence is the momentum flux-profile relationship, which relates the transport of momentum (flux) and the change of wind speed with altitude (profile). This quantity can be computed following the Monin-Obukhov Similarity Theory (Obukhov 1971). However, this theory requires parameters that are hard to measure, both in real life and in the simulations, is only applicable in a restricted range of conditions, and produces predictions that are accurate only up to 20-30% (Foken 2006).The goal of this thesis is to compute the momentum flux-profile relationship using only macro weather parameters, which are readily available in climate simulations; this is done using Data Mining techniques on 17 years of weather data collected from the Cabauw meteorological tower in the Netherlands. Moreover, we asses the impact of different sets of features on the prediction error.Results show that even the simplest linear models are able to compete with the similarity theory, and complex methods such as gradient boosted trees can reduce the mean squared error by almost 50%. Furthermore, these methods are applicable to a much wider range of conditions compared to the similarity theory, while providingroughly the same predictive performance achieved by this theory in its validity range. These results are obtained using wind speed and air temperature at different levels, the soil temperature, and the net radiation at the surface; the improvement offered by the heat fluxes is significant, but of low magnitude. The soil heat flux, the dew point, and the hourly trend of the features do not have a tangible impact on the performance.