Calculating the turbulent fluxes in the atmospheric surface layer using feedforward networks.

Artificial neural networks (ANNs) are common to solve non-linear relationshipsbetween input and output variables as universal approximators. Functional relationscan often be expressed by shallow and quite simple feedforward networks, alsoknown as multi-layer perceptrons (MLPs). This kind of networks was utilised toestimate the turbulent fluxes of momentum and heat in the atmospheric surfacelayer to enhance the modelling of atmosphere and surface interactions. During thetraining and validation procedure about 390,000 data tuples were processed. For eachdata tuple, measurements of temperature and wind speed in two different heightsabove ground and for validation the scales of wind and temperature u∗ and T∗ wererequired. This data originates from hourly multi-year time series from seven differentmeasurement sites covering three various land usage types namely grassland, forest andwetland.Different network set-ups and combinations of input variables were investigated. Allresults were compared to a conventional iterative calculation method following theMonin-Obukhov similarity theory (MOST). Results of best ANNs are comparable to thosecalculated with the classical MOST. These results and the effects of input dataquality, data preprocessing and network architecture will be presented for selectedMLPs. For example, experiments showed that the use of MPLs with only one singlehidden layer is mostly sufficient to describe the variations of the turbulent fluxes. [ABSTRACT FROM AUTHOR]