An Investigation into the Importance of Meteorology in Determining Surface Ozone Concentrations - a Neural Network Approach

Matthew W. Gardner
School of Environmental Science, University of East Anglia, Norwich, Norfolk NR4 7TJ, United Kingdom

Surface ozone concentrations are determined by a complex interaction between radiative chemical and meteorological processes. Ozone is classified as a secondary pollutant since it is not released directly into the atmosphere, but is photochemically produced in the presence of its precursors and sunlight. Ozone concentrations are strongly linked to the meteorological conditions in the boundary layer and also to long-range transport of both ozone and its precursors. Land-sea breezes also influence ozone concentrations at coastal sites.

Much work has been carried out looking at the relationship between meteorology and ozone concentrations on a year to year basis. A simple wind index has been shown to correlate extremely well with the mean summer surface ozone concentration and can account for 64% of its variability (Davies et al., 1992). Feister & Balzer (1991) carried out a regression analysis between daily summer ozone concentrations and 313 meteorological predictors. They showed that 49% of the variability in summer ozone concentrations was related to meteorology. Studies of winter ozone concentrations are less common and tend to be concerned with the effect of intrusions of stratospheric ozone into the troposphere.

To date there exists a discrepancy between the timescale with which ozone chemistry is known to operate and the study of the meteorology-ozone relationship. Ozone chemistry is measured and modelled on timescales of minutes to hours whereas the majority of work investigating the effect of meteorology on surface ozone concentrations has been undertaken on timescales of days to years.

The objective of this research was to investigate the effect of local meteorology on surface ozone concentrations on an hourly basis for a period of at least one year. A feed-forward back propagation neural network was developed to model hourly ozone concentrations from simple meteorological data. The data are from Weybourne, a coastal site in North Norfolk; and include hourly observations of temperature, humidity, irradiance, wind speed, direction and ozone concentrations for the whole year of 1994. No chemical data was used as input to the model. By using purely meteorological input data the degree of ozone concentration variability resulting from changes in weather conditions could be assessed. Any remaining variability could then be attributed to other causes such as the chemical interaction between hydrocarbons and the oxides of nitrogen. Previous work by Boznar et al. (1992) has already illustrated the successful use of a neural network to predict SO2 concentrations given both meteorological and chemical data.

A neural network was used as a tool to carry out this work since the complex relationship between ozone concentrations and meteorology is highly non-linear. It was hoped that the neural network would pick up these non-linear features of the relationship which a conventional statistical technique, such as regression analysis, might overlook. Much effort went into assessing the extent to which the training data and the network architecture affected the overall performance of the network. The model has shown that over a period of a year 48% of the ozone variation can be attributed to changing meteorological conditions.

Further work is currently being carried out to look at the seasonality of the relationship as well as attempting to identify periods where the network did not manage to model the ozone concentration. As a comparison to the neural network technique a conventional multi-variate analysis will be carried out. It is also hoped to extend the database to include 1995. This will enable an examination of any inter-annual variability in the ozone-meteorology relationship.

Davies, T.D., Kelly, P.M., Low, P.S. and Pierce, C.E. 1992. "Surface Ozone Concentrations in Europe: Links With the Regional Scale Atmospheric Circulation", Journal of Geophysical Research, 97, 9819-9832.

Boznar, M., Lesjak, M. and Mlakar, P. 1993. "A neural network-based method for short-term predictions of ambient SO2 concentration in highly polluted industrial areas of complex terrain", Atmospheric Environment, 27B, 221-230.

Feister, U. and Balzer, K. 1991. Surface ozone and Meteorological Predictors on a sub-regional scale", Atmospheric Environment, 25A, 1781-1790.