At present, because the blending coal was taken in some power stations as the major fuel which has too complex physical and chemical characters to build accurate normal near infrared quantitative models in some cases, which brought difficulties for on-line electric coal calorific value detection. For this reason, it was carefully studied that the time domain and frequency domain properties of the power generation coal near infrared spectra, and was proposed that a new quantitative near infrared method named frequency domain self-adaption analysis. The first step, time domain near infrared spectra are converted into frequency domain near infrared signal by Fast Fourier Transform; The second step, the suitable frequency information range by means of valid spectra energy parameter ηE was obtained by this method; The third step, it was constructed that an information volume parameter which is formed by correlation coefficient, standard deviation spectra and coordinate of harmonic in frequency domain to initialize the regression model input parameters’ position; Finally, the optimal model is established by way of discrete frequency domain scooping and synthesized performance function. At the same time, compared with the principle component regression, partial least squares regression, back propagation artificial network, support vector regression and partial least squares regression optimized by genetic algorithm models, it is acquired that a more accurate method which can effectively avoid over fitting and virtual effective models and has a very useful application prospect by verifying the electric coal calorific value. Additionally, this method can be used in other quantitative spectra analysis.