Prediction of ground level ozone concentration using artificial neural network modeling

Abstract

Ozone is a toxic and reactive air pollutant that is hazardous to human health and is detrimental to many living organisms and inorganic substances. Ozone is a secondary pollutant that depends on other air pollutants and meteorological conditions. Prediction of anticipated ozone concentrations is very important for an effective air quality management. Timely warnings of unhealthy ozone concentrations can reduce the associated risks to human health and to sensitive ecosystems. The standard for maximum ozone concentration for hourly exposure is 0.120 parts per million. Recently, a new 8-hour standard for ozone concentrations of 0.080 parts per million was set by the US Environmental Protection Agency. A variety of physically based and statistical models have been developed and applied for predictions of ozone concentrations. An improvement to the statistical models is the application of the Artificial Neural Network (ANN) technology. ANN models can handle complex nonlinear relationships and be easily self-retrained, as new data become available. For this study, the MATLAB - Neural Network Toolbox was used to model the daily ozone levels in Palm Beach County, Florida. After several different trials, the particular model selected and used was a back-propagation one-hidden layer ANN model. The model was tested using different sets of input data including: atmospheric pressure, air temperature, dew point temperature, wind direction, wind speed, and ozone concentration during the previous time step. Temperature, dew point, wind direction, wind speed were provided at five different atmospheric elevations. The maximum number of input parameters included in the trials was twenty-six. In spite of the episodic nature and low frequency of ozone-standard exceeding events, the model was able to predict the ozone fluctuations in a very effective manner.