Application of Neural Networks in Stratified Flow Stability Analysis

Abstract

A feed-forward back-propagation-type neural network was used to predict the flow conditions when interfacial mixing in stratified estuaries commences. This was achieved by training the network to extrapolate data from laboratory experiments performed over many years by several researchers. Before this training could be carried out, however, many decisions concerning the size of the network required and its training parameters had to be made. These decisions were made on the basis of successfully training a similar stratified flow condition, that of thermal wedges downstream of a power plant's outlet, where the theoretical solution is known. Finally, these results were compared with an approximate stability equation utilizing results from inviscid flow theory, rough turbulent flow theory, and laboratory experiments on interfacial friction. Although the agreement was not exact, it was close enough to predict what the stability conditions in real estuaries should be. This prediction was verified with the only prototype data available, that from three fjords, which agreed with both the neural network and theoretical results.