One-step forecasting of seismograms using multi-layer perceptrons

Research in earthquake forecasting has been growing in recent years. Several techniques have been applied: analysis of satellite images for tectonic faults detection, measurement of electro-telluric variations and radon gas levels, data transformations to audible representations of seismograms, to name a few. Artificial neural networks (ANN) have been used in some of these approaches, but there is scarce research related to their applications as function approximation models of the dynamics involved in earth tremor. This type of modeling is useful in these problems because a suitable function approximator can be used as a one-point predictor. In this work we investigate the capability of multi-layer perceptrons (MLP) to model data from earth tremor, building one-point predictors trained over seismograms taken from the earthquake occurred at Mexico City of September 19, 1985. MLP´s were trained using the resilient backpropagation (PROP) and the Levenberg-Marquardt (LV-MQ) algorithms, obtaining in average a MSE = 0.7964 in the first case and 0.7646 for the second case. This results show the ability of MLP´s to model this kind of non-linear data for one-point prediction.