Spectral detection based on spectrophotometry is an important multi-component concentration detection method. At present, commonly used machine learning methods in the field of spectral analysis can only be used for prediction and cannot analyze how the concentration of each component affects the spectrum. In addition, for common spectral parallel drift in spectrophotometry, traditional derivative preprocessing methods are susceptible to noise and cannot reverse restore the original spectrum. These issues are not conducive to obtaining accurate interpretable spectral models and limits further improvement in detection accuracy. To solve the above problems, we applied mathematical methods to establish a basic model of absorbance spectra for multi-component mixed solutions, and analyzed the influence of spectral parallel drift on it. Then, we proposed an anti-drift modeling method based on the adjacent difference method. This method not only eliminates drift in the spectrum, but also achieves reverse optimization estimation of the original spectral model. In addition, in response to the phenomenon that the adjacent difference method amplifies spectral noise, we analyzed its reasons and proposed a disorderly difference method that can perfectly balance the advantages of adjacent difference method and span difference method. This method has better performance when applied to anti-drift modeling. Finally, the effectiveness of the method and improvement measures was validated by applying them on cobalt ion spectral dataset with different concentrations of iron and copper ions in zinc hydrometallurgy.