Data mining to characterize ozone behavior in Baltimore and Washington, DC

Data mining exercises were used to describe long-term ambient ozone concentrations in Baltimore and Washington based on meteorological conditions as well as upwind and previous day concentrations. Ozone production occurs mainly during the summer months and is influenced by a variety of meteorological parameters. Fifteen years of daily ozone measurements (May–September) were subset into five daily meteorological clusters using an expectation–maximization technique (based on available temperature, solar radiation, precipitation, and other meteorological parameters). Rule association and classifier models were used to quantify the contributions from high overnight ozone concentrations at upwind rural sites (21–36 ppb). Data clustering followed by further subsetting was used to examine the effects of low level jets (on average 5–7 ppb higher when jets were observed).To interpret interannual ozone trends, the year-to-year fluctuations in summertime meteorology must be accounted for and adjusted. The historical trends of each meteorological cluster were measured to determine which trends were statistically significant. By separately analyzing the trends of each meteorological cluster, the results were weather-normalized to discount the effects of rainy or drought years in the analyses.Seven different ozone parameters were investigated, and three of the five clusters showed weak or no trends for the time periods 1991–2004 and 1999–2004. However, a Baltimore Cluster which represents sunny conditions with variable winds and a high temperature difference between morning surface and aloft temperatures showed statistically significant decreases in three ozone parameters from 1991 through 2004. The most compelling evidence came from the cluster representing sunny and hot conditions with high wind speeds from the west and northwest and showing statistically significant decreasing trends for six of the seven ozone parameters from 1999 through 2004.