It's common to use the method of continuous spectroscopy in water quality testing. But there're some problems with it. For example, the scanning results have a large number of nonlinear signals, and the covariance between variables is serious, which can lead to a decrease in the model prediction accuracy. In this paper, the standard solutions of nitrate nitrogen (NO3-N) and nitrite nitrogen (NO2-N) were used as the subject to be tested, and the data of the scanned waves and absorbance were obtained by use of spectral detector. The data were processed by noise reduction first and then the random forest (RF) algorithm was adopted to establish the regression relationship between concentration and absorbance. For comparison, partial least squares (PLS) and support vector machine (SVM) algorithm models were also established. For the same given data, the three reverse models can make the projection of the concentration respectively. The experimental results show that the RF algorithm predicts NO2-N concentrations significantly better than the SVM algorithm and PLS algorithm. This proves that the RF algorithm has good prediction ability in spectral water quality detection because of its high model accuracy and better adaptability, which could be a reference for similar research on continuous spectral water quality online detection.