Speaker
Description
The quasi-biennial oscillation (QBO) represents a significant component of atmospheric dynamic variability in the stratosphere. Disruptions in the regular, if complex, nature of the QBO observed beginning in 2015 have reduced the power of traditional modeling approaches using linear dimensionality reduction techniques, e.g. principal component analysis (PCA). While the use of additional principal components in trend analyses can increase efficacy, shortcomings may be present in the traditional approach given the unprecedented behavior observed in the last decade. A nonlinear PCA (NLPCA) approach is an option for representing multidimensional variance in nonlinear regimes while retaining the ability to extract linear and non-linear structures from data. NLPCA can potentially represent nonlinear perturbations present in the QBO and adequately account for resulting atmospheric impacts. The method presented implements an auto-associative neural network (AANN) which is a multi-layer feed-forward network. The AANN is trained by adjusting neuron weight (coefficient) values to minimize the loss between the original dataset and the reproduction using the limited number of neurons in a series representing encoding, compressive, and decoding layers. A python package has been created to implement NLPCA using an AANN with applicability geared toward atmospheric datasets and data structures. Assessments are performed using models of ozone concentrations, and comparisons between linear and nonlinear PCA coefficient QBO representations is performed using seasonal time series analysis and trending techniques.
Keywords: nonlinear principal component analysis, quasi-biennial oscillation, auto-associative neural network, ozone
