Simulating Pressure And Velocity Time Series With Artificial Neural Networks: Some Advantages And Pitfalls

Abstract

Three examples of time series simulations of pressure and velocity fluctuations using artificial neural networks were discussed: (i) a spatial interpolation of pressure time series on the roof of a low building in a thick, turbulent boundary layer, (ii) a simulation of two velocity components at multiple spatial locations simultaneously in the turbulent far wake of a circular cylinder, and (iii) a simulation of pressure time series around the surface of a circular cylinder in a crossflow. For the spatial interpolation a backpropagation network was used, while for the other two simulations, the fuzzy ARTMAP neural classifier was used. It was shown that the fuzzy ARTMAP captured the energy of the fluctuations over a wider range of scales than the backpropagation network because of its architecture, even though the input and output types were similar. The fuzzy ARTMAP is based on a clustering-type of pattern recognition while the backpropagation network is more deterministic, i.e., more like an empirical curve-fit to the data. This appears to allow the fuzzy ARTMAP to capture the dynamics of the flow field to a greater extent.