The fault diagnosis problem for dynamic power systems is treated, the nonlinear dynamic model based on a differential algebraic equations is transformed with reduced index to a simple dynamic model. Two nonlinear observers are used for generating the fault signals for comparison purposes, one of them being an extended Kalman estimator and the other a new extended Kalman filter with moving horizon with a study of convergence based on the choice of matrix of covariance of the noises of system and measurements. The paper illustrates a simulation study applied on IEEE 3 buses test system.