Neural Network Based Air Quality Data Filling
Free (open access)
G. Latini, G. Passerini & S. Tascini
Neural Networks (NN) have become a fundamental tool among data-handling procedures and even more concerning environmental data. In this work we present an application of neural networks to air quality data prediction. Both primary pollutants (mainly SO2, CO) and photochemical pollutants (particularly ozone) have been considered but the focus has been set on statistical correlation between precursors and secondary pollutants. After a preliminary study of the phenomena, the work consisted in the following steps: NN architecture choice (we considered Multi-layer Perceptron Networks, recurrent networks and Self Organising Networks), NN set-up, and input handling. Ozone precursors (e.g. NOx) and meteorological variables have been considered (solar radiation, wind velocity and temperature), noticing that only non-linear relationships were present. We performed an input correlation analysis and we considered normalisation processes and post-training analysis. For the NN training we selected the most representative periods regarding ozone cycle. The final step was the network validation: generalisation capability and prognosis of never processed data-set have been verified. To maximise the process automation, a software tool has been implemented in the MatlabTM environment. The NN validation showed encouraging results and we successfully extended the SW tool application to the air quality data filling. 1 Introduction Time series are indispensable elements among the air quality study as well as in many other research fields in which reliable data are necessary for modeling and validation. Series length and completeness are fundamental requirements .