The temperature error characteristics of the fiber optic current sensor (FOCS) based on polarization modulation principle is theoretically analyzed to meet the requirement of engineering application. The temperature compensation of FOCS is realized by least squares method and BP neural network respectively. The nonlinear temperature error correction is realized. And the experimental results of the two temperature compensation algorithms are compared and analyzed. The result shows that the temperature compensation result based on neural network algorithm is better than that of least squares method. The full temperature experiment of FOCS is used to verify the reproducibility. The experimental results show that after corrected by the neural network algorithm, the temperature error of FOCS in the range of from -5 ℃ to +50 ℃ is less than 0.5%.