Wind dispersion of carbon dioxide leaking from underground sequestration, and outlier detection in eddy covariance data using extreme value theory.

摘要

The first part of this thesis describes how a relatively simple model can be used to predict the dispersion by turbulent winds of CO 2 leaking from underground sequestration. It is often a good approximation to model both the wind velocity and the turbulent diffusivity by power-law functions of height, which allows an analytical solution to the advection-diffusion equation. This solution is compared to a coupled simulation of CO2 leakage which included both aboveground and belowground layers in the same code. Since the time scale is much faster in the air than underground, the two domains were not coupled in practice. Thus it is more efficient to use the analytic solution in the aboveground domain, and it also provides more realistic boundary conditions.;The second part of the thesis concerns measurements of the flux of trace gases between ecosystems and the atmosphere by eddy covariance. There is a quality control issue with "spikes" from the sonic anemometer; since these spikes occur together in both wind and temperature channels, they can introduce significant errors in the measurement of heat flux. Common methods of spike detection rely on excluding points which exceed some number of standard deviations from the mean. However, the probability of large deviations in valid data is not known in advance, and varies between records. A new algorithm is developed using extreme value theory to identify points which are very unlikely to be generated by the same process as the rest of the data. The algorithm starts from the center of the distribution and moves up toward the maximum, using the "spacing theorem" to predict the expected spacings between successively ranked points, until a gap is found whose probability is below the desired significance level; all points beyond the gap are classified as outliers. The expected gap is predicted based only on points which have already been judged valid, thus avoiding any need for repeated passes after rejecting points. The spacing method is computationally efficient, and can preserve valid data in samples where valid points can be as much as 10 to 15 standard deviations from the mean.