The time-varying nature of electric arc furnace (EAF) gives rise to voltage fluctuations, which produce the effect known as flicker. The ability of static VAr compensator (SVC), a widely used method for flicker reduction, is limited by delays in reactive power measurements and thyristor ignition. To improve the SVC performance in flicker compensation, arc furnace reactive power can be predicted for a half cycle ahead by using a suitable dynamic model for EAF. To obtain suitable models for EAFs reactive power, this thesis uses huge field data including voltages and currents for eight EAFs installed in Mobarakeh Steel Industry. The reactive power times series are produced from the recorded currents and voltages based on the fundamental reactive power. A novel and comprehensive study including linear analysis, online calculation of models coefficients and non linear analysis is performed for dynamic modeling of arc furnace reactive power variation. In the linear analysis, ARMA models are used for reactive power time series and, appropriate orders of ARMA models are obtained by performing various statistical tests. The results show that reactive power times series with 10 sec length are stationary and can be suitably modeled by ARMA(2,1), ARMA(2,2) and ARMA(3,2) models. In this analysis vector ARMA models are also investigated. The better performance of the compensator in the case of employing predicted reactive power of EAF is demonstrated rather than that of the conventional method by using three new indices that are defined based on concepts of flicker frequencies and the power spectral density. It is shown that the EAF reactive power models coefficients and their variations are different from one data record to another and do not follow any specific law. Therefore, it is necessary to update the model coefficients for prediction purposes. For this purpose, the transient and steady state performances of NLMS, RLS and online genetic methods are studied based of new indices. In the nonlinear analyses, non linear parameters of the reactive power time series are obtained. Then by using different tests, it is demonstrated that the percent of time series that have nonlinear properties is between 20% to 60% for different tests. In addition by study of maximum lyapanov exponent, it is found that reactive power time series are not chaotic process. Finally by using a new method based on the residual time series and some new indices, it is demonstrated that the ratio of time series nonlinear deterministic