The laser with a wavelength of 1570 nm is used to analyze the hydrogen sulfide gas in the background of natural gas. The gas mixture containing the hydrogen sulfide with a volume fraction of 0-10-4 is produced in an automatic gas mixing station, and 92 groups of the spectral data with a stable state are obtained. The regression model of extreme learning machine (ELM) is adopted for the inversion calculation of the concentration of hydrogen sulfide. The nonlinear iteration partial least square (NIPALS) algorithm is introduced into the spectral pretreatment. The ELM regression model is established by using the spectral feature vector and the concentration vector, and is evaluated by the five-fold cross validation method. The test results show that, the regression predicating accuracy of the spectral data obtained by feature extraction is improved by 25% than that by direct ELM, and the model operation time is reduced from 0.12 s to less than 10 ms. The spectral pretreatment by the feature extraction can reduce the training time of ELM model and can also improve the analysis accuracy and the real-time capability of the analyzers.