Neural network models unlike traditional multiple linear regression

Fig. 5. Distributions of individual aerosol factors' modeled daily O3 impact at LAVO and GRBA.Figure optionsDownload full-size imageDownload as PowerPoint slide
The inability of the models to simulate some of the lowest O3 concentrations, particularly in winter (Fig. 3, b and c), contributes to the narrower SD ranges. While low O3 is not of regulatory interest, improving the meteorological surrogates to capture the effects precipitation or snow cover appears warranted in future work.
Some negative numbers may also be a computational result of using the 24-hour IMPROVE samples. The day-night Forskolin in these samples adds some uncertainty to the presumption of a uniform air mass represented by a single sample, a simplification forced on our model by the IMPROVE protocol. Future work with longer sets of time-resolved aerosol data is needed to clarify this issue.
3.3.2. Neural net O3 simulations compared to regional O3 modeling
Fig. 6 compares the neural net O3 simulations performance statistics to results reported from high resolution (12-km and 36-km nested-grids) regional modeling of 8-hour O3 at CASTNET sites in the Western US for 2008 (WRAP, 2013). The neural net simulations appear to perform about as well as highly refined regional models for these remote sites.