Self-organizing feature maps predicting sea levels

Abstract

In this paper, a new method for predicting sea levels employing self-organizing feature maps is introduced. For that purpose the maps are transformed from an unsupervised learning procedure to a supervised one. Two concepts, originally developed to solve the problems of convergence of other network types, are proposed to be applied to Kohonen networks: a functional relationship between the number of neurons and the number of learning examples and a criterion to break off learning. The latter one can be shown to be conform with the process of self-organization by using U-matrices for visualization of the learning procedure. The predictions made using these neural models are compared for accuracy with observations and with the prognoses prepared using six models: two hydrodynamic models, a statistical model, a nearest neighbor model, the persistence model, and the verbal forecasts that are broadcast and kept on record by the Sea Level Forecast Service of the Federal Maritime and Hydrography Agency (BSH) in Hamburg. Before training the maps, the meteorological and oceanographic situation has to be condensed as well as possible, and the weight and learning vectors have to be made as small as possible. The selforganizing feature maps predict sea levels better than all six models of compari