For the complex nonlinear relationship between zero bias and temperature of fiber optic gyro (FOG), extreme learning machines (ELM) model is introduced to compensate the zero bias temperature error of FOG. Aiming at problems of insufficient prediction accuracy and stability of a single ELM and its sensitivity to singular samples, the adaptive boosting (AdaBoost) algorithm is introduced to establish an ELM-AdaBoost prediction model to improve the FOG performance. The temperature error mechanism of fiber optic gyroscope and the influence of model parameters on the prediction accuracy are analyzed, and the determination methods of the number of neurons in the hidden layer of ELM algorithm and the number of iterations of AdaBoost algorithm are given. The simulation results show that. The compensation effect based on ELM-AdaBoost prediction model is better than the multiple linear regression model and single ELM neural network model, and has good generalization performance and temperature applicability, and the root mean square error of gyro zero bias is reduced by more than 93% after compensation, which significantly improves the stability performance of fiber optic gyro zero bias.