Abstract: Metallized film capacitor (MFC) plays a great role in the traction system of rail transit. The accurate prediction of its lifetime is helpful to optimize the operation and maintenance strategy of the converter system, and reduce the cost of system operation and the losses caused by failures and accidents. Aging test requires a lot of cost and time, which leads to the problems of small number of data samples and few parameter measurement points in current lifetime prediction. Hence, this paper took the advantages of Support Vector Regression (SVR) that could be suitable for these samples and implements Deep Belief Network (DBN) to in-depth extract features from time series without sophisticated and unknown empirical degradation function. A model which combines DBN and Binary SVR (BSVR) was built, where one with linear kernel function performs Recursive Multi-step Forecast Strategy (RMFS) on degradation series and another based on the Gauss kernel function completes the error compensation training after obtaining the error, which eliminates accumulated error of RMFS. Thereupon, a residual lifetime prediction method based on deep feature extraction and error compensation was simulated on the dataset of capacitance degradation of MFC under aging test. The experimental results showed the average prediction error at different prediction starting points was only 7.08%. Compared with existing methods, the proposed method greatly heightens the residual lifetime prediction accuracy of MFC and enhances the reliability as well as generalization ability.