The near-infrared spectrum of some common solvents, such as water, has very high sensitivity to temperature. So, the effect of temperature can not be ignored in NIRS. When temperature changes, the transmission spectra of the solution will also change. The influence of the temperature was deduced theoretically based on the Lambert-Beer’s law in the present paper. And a method for temperature correction of sample’s spectra was proposed. By using the change in pure solvent’s absorbency with temperature disturbance, the method is used to correct the spectra of prediction samples. The spectra of glucose aqueous solution and albumin aqueous solution were measured at different temperatures. The calibration models of glucose and albumin concentration prediction were built respectively by using the calibration sample’s spectra at 30 ℃. The absorbency of pure water has different value at different temperature, and this difference was used to correct the spectra of sample which was measured at temperatures other than 30 ℃. The experiment results showed that the curves of absorbancy difference of prediction samples, which were measured at different temperatures, show good superposition after spectra correction. And the root mean square error (RMSEP) of prediction was reduced obviously. It also means that the influence of temperature on the spectra can be eliminated effectively after spectra correction by using the method proposed in this paper.