In this study, particulate matter (PM), total suspended particulate (TSP), PM10, and PM2.5 fractions) concentrations were recorded in various cities from south of Romania to build the corresponding time series for various intervals. First, the time series of each pollutant were used as inputs in various configurations of feed-forward neural networks (FANN) to find the most suitable network architecture to the PM specificity. The outputs were evaluated using mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and Pearson correlation coefficient (r) between observed series and output series. Second, each time series was decomposed using Daubechies wavelets of third order into its corresponding components. Each decomposed component of a PM time series was used as input in the optimal feed-forward neural networks (FANN) architecture established in the first step. The output of each component was re-included to form the modeled series of the original pollutant time series.
Part of the book: Artificial Neural Networks