This paper is concerned with modelling and simulation of near-source earthquake ground motion. Recent studies have revealed that these motions show heavy non-stationary behaviour with very low frequencies dominating parts of the earthquake sequence. Modeling and simulation of this behaviour is of crucial importance in the design of flexible structures and other applications. This paper examines three approaches for the modelling and simulation of non-stationary nearsource ground accelerograms: The first one makes use of
ARMA models combined with frequency and variance "stabilization". The second is based upon inherently nonstationary Time-dependent ARMA (TARMA) models, the parameters and variance of which are allowed to be explicit functions of time. The third approach is based on Neural Networks. The three approaches are used for the modelling and simulation of an accelerogram characterized by an epicentral distance of 16 km and measured during the 1979 Imperial Valley earthquake in California (U .S .A.). The results of the study indicate that while all three approaches can successfully predict near-source ground motions, the Neural Network based
one gives somewhat poorer simulation results.