According to real-time and precision requirements of online compensation, a support vector regression method improved by particle swarm optimization was proposed to establish the fiber optic gyroscope temperature error compensation model. A real-time acquisition method of temperature change rate based on multi-data window was adopted to meet the requirements of online compensation and model input. The fiber optic gyroscope was placed in a thermostat for temperature change test range from -15 to 50℃ to obtain the measured data. The temperature and temperature change rate were taken as inputs, and the least squares, radial basis function neural network and particle swarm optimization support vector regression algorithm were modeling respectively. The results show that the proposed model achieve the best compensation effect. The real-time compensation comparative experiment verified that the proposed model has good real-time compensation performance and generalization ability for non-training data.