In order to improve the measurement accuracy of the embedded time grating angular displacement sensor, starting from the formation mechanism of the sensing signal, the cause of the short period error is analyzed in detail. Through the winding equivalent analysis and excitation signal analysis, it is determined that the main characteristics of the short period error are the first order error and quadric error. The source of the first order error is the zero residual error and direct current component error, and the source of the quadric error is the excitation signal quadrature error. Aiming at the compensation of the short period error, an error compensation method based on extreme learning machine is proposed. The model optimal parameters are obtained through training the measured values and real sample values. According to the model parameters, the short period error model is established, which is used to realize the short period error compensation. The experiment result shows that the analysis result of short period error is consistent with the actual characteristic of the sensor error, and using the proposed error compensation method the short period error of the sensor is reduced by about 96%, which is a greatly reduction. The comparison and repetitive experiments show that the accuracy of the proposed method is improved doubly compared with that of the harmonic compensation method, the error compensation effect is superior. Besides, the proposed method possesses good measurement stability, which has important theoretical and practical significance for improving the measurement accuracy of the embedded time grating angular displacement sensor.