1.3
Artificial Neural Networks
Although the scientific basis of computational models allows their results to
reveal useful information about their causes and the processes involved, they
can be computationally- and temporally-intensive when modeling fine-scale
7
processes over a large area. Another modeling style is based on statistical
analysis; put simply, the idea is to to use past behavior, rather than scientific
theory, to predict future behavior. A dataset including both the input factors
and the resultant air pollution is fed into a computer model, which refines
itself on the patterns in that dataset. The resultant model can then be used
to predict the pollution levels for new input values.
Neural networks are well-established as a modeling technique for air pollution; in 1991, Boznar et al. [7] used a neural network to model SO 2 emissions
from a power plant. Since then, they have been used to predict a wide range
of pollutants. Comrie [14] found that a neural network model predicted ozone
to within 1-2 ppb of observed values in the 40-70 ppb range, a slightly better performance than a multiple regression model applied to the same data.
Gardner and Dorling [22] ran multiple neural network models predicted NOX
concentrations in London with different timesteps, with their hourly prediction model explaining 92% of observed variation. More recently, Cai et
al. [9] created an ANN model that predicted the concentrations of multiple
pollutants with correlation coefficients over .85, higher than a simple source
dispersion model [24].
Those models, like most neural network pollution models, were based
solely on meteorological and pollution data. Some recent work has incorporated inputs representing other sources and sinks. Zheng et al. [54] include a
large variety of data sources, including taxi movement, human mobility, and
city points of interest, in an ANN model of Beijing'pollution. Though theii
model produced accurate predictions, it did so for air quality index categories
rather than an absolute, numeric measurement of pollution. It also was based
on a 1km-square grid; while this provides a much denser picture of urban air
pollution than the earlier studies, which implemented a model for only a few
points around a city, it fails to capture finer variations in pollution.