US EPA

4977272 

Technical Report 

Data Mining to Determine Local Effects of Mercury Emissions 

Walsh, K; Milligan, M 

2005 

GRA and I 

e 1 

Tools for knowledge discovery through databases (KDD) can provide air quality planners with descriptions of long-term behavior at monitor sites. In this project, the Maryland Power Plant Research Program used the Cubist (trademark) program to construct a piecewise linear regression model to describe mercury deposition rates and concentrations at two sites in terms of hourly power plant heat input (a surrogate for emissions) and metereorological variables. The ambient mercury measurements represented more than five years of data at the Marcell, Minnesota and Chassahowitzka, Florida sites of the Mercury Deposition Network. The daily heat input data represented two power plants (Boswell Energy Center and Florida Power's Crystal River Energy Complex) that are predominantly coal-fired. The advantages of local versus regional controls for mercury emissions have recently been debated. This work aimed to identify the meteorological conditions that lead to an effect of local emissions on local ambient mercury conditions. The long time series of ambient mercury deposition data at the Marcell and Chassahowitzka sites were compared with the simultaneous heat input quantities at the two nearby power plants. Meteorological conditions were included in the data sets, and the Cubist software generated piecewise linear regression fist that describe the deposition data based on heat input and meteorological parameters. The software created models of several linear expressions that apply under particular meteorological conditions and sometimes overlap. The investigation showed that the effect of power plants on nearby mercury deposition sites only manifested itself under particular meteorological conditions. The data analyses and findings are discussed in detail.