Analysis of air pollution data at a mixed source location using boosted regression trees

research

air quality

source apportionment

meteorological normalisation

data analysis

Boosted regression trees are applied to air pollution data at a mixed-source location to disentangle the contributions of different emission sources and meteorological influences, demonstrating a flexible machine learning approach for complex air quality data.

Authors

D.C. Carslaw

P.J. Taylor

Published

January 1, 2009

Analysis of air pollution data at a mixed source location using boosted regression trees

D.C. Carslaw, P.J. Taylor

Atmospheric Environment, 2009

This paper explores the use of boosted regression trees to draw inferences concerning the source characteristics at a location of high source complexity. Models are developed for hourly concentrations of nitrogen oxides (NO_x) close to a large international airport. Model development is discussed and methods to quantify model uncertainties developed. It is shown that good explanatory models can be developed and further, allowing for interactions between model variables significantly improves the model fits compared with non-interacting models. Methods are used to determine which variables exert most influence over predicted concentrations and to explore the NO_x dependency for each. Model predictions are used to estimate aircraft take-off contributions to total concentrations of NO_x and determine how these predictions are affected by annual variations in meteorological conditions and runway use patterns. Furthermore, the results relating to the aircraft contributions to total NO_x concentration are compared with those from a more detailed independent field campaign. Finally, we find empirical evidence that plumes from larger aircraft disperse more rapidly from the point of release compared with smaller aircraft. The reasons for this behaviour and the implications are discussed.

BibTeX