RESUMO
In this paper, we described generation and performances of feedforward neural network models that could be used for a day ahead predictions of the daily maximum 1-h ozone concentration (1hO3) and 8-h average ozone concentration (8hO3) at one traffic and one background station in the urban area of Novi Sad, Serbia. The six meteorological variables for the day preceding the forecast and forecast day, ozone concentrations in the day preceding the forecast, the number of the day of the year, and the number of the weekday for which ozone prediction was performed were utilized as inputs. The three-layer perceptron neural network models with the best performance were chosen by testing with different numbers of neurons in the hidden layer and different activation functions. The mean bias error, mean absolute error, root mean squared error, correlation coefficient, and index of agreement or Willmott's Index for the validation data for 1hO3 forecasting were 0.005 µg m-3, 12.149 µg m-3, 15.926 µg m-3, 0.988, and 0.950, respectively, for the traffic station (Dnevnik), and - 0.565 µg m-3, 10.101 µg m-3, 12.962 µg m-3, 0.911, and 0.953, respectively, for the background station (Liman). For 8hO3 forecasting, statistical indicators were - 1.126 µg m-3, 10.614 µg m-3, 12.962 µg m-3, 0.910, and 0.948 respectively for the station Dnevnik and - 0.001 µg m-3, 8.574 µg m-3, 10.741 µg m-3, 0.936, and 0.966, respectively, for the station Liman. According to the Kolmogorov-Smirnov test, there is no significant difference between measured and predicted data. Models showed a good performance in forecasting days with the high values over a certain threshold.