Neural Networks for Characterization and Forecasting in the Boundary Layer via Radon Data

The complexity of air-pollution physical-chemical processes in the boundary layer (BL) is well known: see, for instance, Stull (1988) and Seinfeld and Pandis (1998). In this framework, we do not make any attempt at reviewing the manifold use of neural networks (NNs) for air-pollution assessments and forecasting. Instead, we focus just on the (complex) physics of the BL and discuss the coupled use of an original index of the BL properties (radon concentration) and of NN modeling in order to obtain interesting results for characterizing and/or forecasting important variables in the BL, like the concentration of a dangerous primary pollutant (benzene) and the 2-h evolution of stable layer depth. In this scenario, the particular strategies for applying a NN model are described, showing how they lead to important original results, for grasping the BL physical behavior. In doing so, one can discover the usefulness of an empirical AI data-driven approach to investigating a complex system that is very difficult to deal with in terms of dynamical models.

In the next section, a brief introduction to fundamentals of radon detection will be presented and the qualitative and quantitative relevance of radon concentration for summarizing the physical state of the BL will be discussed. In particular, we will present the structure of a box model (based on radon data) for estimating the nocturnal stable layer depth.