A genetically optimised neural network for prediction of maximum hourly PM10 concentration

Concentrations of ambient air particles have been found to be associated with a wide range of effects on human health. PM 10 concentrations are usually used as a standard measure for air pollution. Increase in the level of PM10 has been associated with increases in mortality and cardio respiratory hospitalisations. Therefore, prediction of ambient levels in certain environments is of great importance, especially in urban and industrialised areas. The present work aims to develop an adaptive system based on Artificial Neural Networks (ANN) that will allow the prediction of the maximum 24-h moving average of PM10 concentration. A special ANN architecture is employed, the Time Lagged Feed forward Network (TLFN), with genetically optimised topology and learning parameters. This type of network is able to process information over time and produce time-varying nonlinear mappings from the chosen input variables to the predicted value. The network is trained and testified by hourly data collected at two air pollutant-monitoring stations in an urban and nearby industrial location in northern Greece. The initial study presented in this paper involves a small subset of the available data that were used to validate the approach and the chosen ANN architecture.