Nonlinear and discontinuities modeling of time series using artificial neural network with radial basis function

Abstract

At present, the Global Navigation Satellite System (GNSS) is used to obtain one Geodetic Reference System by achieving a network of a continuous GNSS monitoring. Due to a variety of factors, the time series may expose data gaps and discontinuities that change its trend. This study proposes to use radial basis function of a neural network with the main objective to generate a single modeling that resolves such difficulties. Examinations were realized using data of the time series of the ellipsoidal height of the station CONZ continuous monitoring located in Concepcion-Chile, between GPS the weeks 1170 and 1833. This station has been chosen due to the circumstances that their time series reflect non-linear behavior, the appearance of data gaps, and it demonstrates also changes in its trend as result of an intense earthquake. The Artificial Neural Network (ANN) technique has been involved with the purpose to realize a single model. The structure of the ANN that has been obtained after training was Radial Basis Function (RBF). In that specific RBF the input vector has been in the week GPS and the output vector has been the ellipsoidal height. Finally, the time series was modeled for the determination of the capacity of a potential generalization in such possible prediction. The result of such prediction yields a mean arithmetic 0.01mm, and standard deviation 2.5 mm. This leads to the conclusion that it may be possible to use in any time series one single neural model only to obtain a reasonable prediction.