To address the bias drift and reduced accuracy caused by environmental temperature fluctuations in fiber-optic gyroscopes，this study proposes a temperature error compensation method based on adaptive variational mode decomposition(AVMD)and an improved grey wolf optimizer-extreme learning machine(IGWO-ELM). First，the AVMD algorithm is introduced to adaptively determine the optimal mode number and penalty factor of variational mode decomposition by minimizing the average envelope entropy，thereby adaptively decomposing the temperature signal into trend，fluctuation，and noise components. Simultaneously，the same decomposition parameters are used to decompose the gyroscope output signal，achieving multiscale signal separation. By acquiring temperature and output data of the fiber-optic gyroscopes from . 40℃ to +60℃，IGWO-ELM submodels are established for the trend and fluctuation components of temperature to predict the corresponding components in the gyroscope output. For the noise component，wavelet threshold denoising technology is employed to suppress high-frequency interference. Finally，the prediction results of the three components are reconstructed to obtain the final compensated output. Experimental results demonstrate the effectiveness of this method in separating multiscale features under temperature variations，thereby significantly suppressing bias drift and enhancing the accuracy and stability of the compensated gyroscope.