Aiming to early fault prediction problem of connected-grid PV converters, a fault prediction method is presented based on RLS-SVR in this paper. The degradation parameter couplings of vulnerable components and its effects on conditions of converters and selection of features for fault prediction are analyzed in this paper in system level, and then the prediction method using the relative variables of features as converter states is proposed. In order to reduce the degeneration prediction from the working condition, the method firstly builds the feature fitting model without degeneration with work condition as input and state features as output by using robust least squares support vector regression (RLS-SVR). Then, the time series of relative variable features of converters is obtained by combining the online working conditions time series and feature time series with no degeneration fitting model during the degradation procedure. At last, the prediction model of the relative variable time series of features of converters is built based on the degradation time series and using RLS-SVR. The prediction method is simple, low cost, high precision, and without adding other sensors. Experimental results of single-phase PV connected-grid converter show that the proposed method is feasible and effective.