In order to improve time-grating sensor’s measurement accuracy, according to the characters of errors, targeted modeling thought is proposed for the periodic part and the random part of dynamic measurement errors. Fourier series approach model is presented to establish the periodic error compensation model. By using least square method, parameters of approach model can be determined, the major of which are chosen to realize error separation. In order to model the random part of errors, support vector regression (SVR) model is proposed. The global optimal parameters of the model SVR which are the best kernel function parameter g and penalty factor C are evaluated by using cross validation method. A synthesis error compensation system is investigated, and realize error compensation. The accuracy analysis certificates that time-grating sensor’s dynamic measurement error is reduced from 38.2″to 3″by applying the error compensation method and compensation system. Sensor’s accuracy is effectively improved by using the compensation system.