It is a challenge task to enhance the analysis accuracy of laser-induced breakdown spectroscopy (LIBS) in high temperature applications when certified standard samples used for building calibration curves at high temperature are limited or not available. A novel LIBS quantitative analysis method for alloy steel at high temperature via feature-based transfer learning (FTL) is proposed. The spectral data of calibration samples at room temperature and the spectral data of uncalibrated samples at high temperature are together transferred into a high-dimensional feature space using kernel function mapping where an LIBS regression model is trained and established. For testing samples, the measured spectra at high temperature are mapped into the high-dimensional feature space with the same kernel parameters used in the training process, and then the concentration results can be obtained by the regression model. Experiments on certified alloy steel standard samples were conducted, in which 12 samples with both the concentration information and the measured spectra at room temperature and 8 samples only with the spectra measured at high temperature were used to train the analysis model. The 8 samples at high temperature were used for testing. The experimental results of the Cr concentration showed that with feature-based transfer learning, the mean relative error decreased from 32.31% to 6.08%. The proposed method does not need the element concentration for samples at high temperature to build the regression model, which provides a feasible and effective approach for LIBS analysis of samples at high temperature, such as fast industrial measurements in iron and steel smelting production processes.