In order to discuss the effect of different distribution of components concentration on the accuracy of quantitative spectral analysis, according to the Lambert-Beer law, ideal absorption spectra of samples with three components were established. Gaussian noise was added to the spectra. Correction and prediction models were built by partial least squares regression to reflect the unequal modeling and prediction results between different distributions of components. Results show that, in the case of pure linear absorption, the accuracy of model is related to the distribution of components concentration. Not only to the component we focus on, but also to the non-tested components, the larger covered and more uniform distribution is a significant point of calibration set samples to establish a universal model and provide a satisfactory accuracy. This research supplies a theoretic guidance for reasonable choice of samples with suitable concentration distribution, which enhances the quality of model and reduces the prediction error of the predict set.