Convolutional Neural Network to forecast complex systems: the nitrogen oxides concentration case

In this work, a data-driven forecasting system based on Convolutional Neural Network is formalized and applied. Convolutional Neural Networks allow the extraction of features from an image or a series of images. In this context, these models are applied to reproduce the concentration level of nitrogen dioxide in advance, starting from the emission maps of nitrogen oxides in the atmosphere. The system is applied to the Milan municipality (Italy), an area often affected by high levels of pollution, with very good performances in terms of both statistical analysis (errors and correlation) and threshold indexes (hit ratio and false alarm ratio). In the near future, the model will be included into a model predictive control system to manage air quality in the area.