Polycyclic aromatic hydrocarbons (PAHs) have strong carcinogenicity and threaten human health. In the complex water quality detection environment, when the concentrations of PAHs are detected by fluorescence spectrum, the spectral signal may contain obvious non-stationary noise due to the influence of Rayleigh scattering in the measured spectrum. The common multiple sampling and averaging method tend to generate obvious measurement error in the PAHs spectrum, thereby leading to low detection accuracy of PAHs. In this paper, an optimal detection method for PAHs concentration based on three-dimensional (3D) fluorescence spectral analysis and N-way partial least square (N-PLS) is proposed. First, the spectral features of the four PAHs solutions of phenanthrene, fluorene, acenaphthene and fluoranthene were analyzed. The Rayleigh scattering noise in the spectrum was eliminated by fitting the scattering band data point values while the original spectrum information was preserved as much as possible. The features such as the mean, variance, and one-dimensional marginal distribution of the four PAHs spectrum were extracted,and the similar spectral data samples were merged according to samples classification of four spectral data by feature clustering analysis. Secondly, the N-PLS model was established based on the relationship between the spectral signal of the correction set and the different PAHs concentration. Subsequently the N-PLS model was used to predict and analyze the concentration of various PAHs, and verify the relationship between the PAHs concentration and the fluorescence intensity of the spectral data. Finally, the concentration residuals were modified by bilinear decomposition, the concentration residuals between aqueous solutions containing various PAHs and real water samples were verified, and the prediction errors of PAHs under different parameters were also analyzed. The experimental results showed that the phenanthrene solvent exists two obvious fluorescence peaks, and their excitation and emission wavelengths are 285/245 and 315/345 nm respectively. Both fluorene and fluoranthene have six obvious fluorescence feature peaks. Their excitation and emission wavelengths are 265/255, 325/345, 335/325, 365/355 nm, 385/395 and 405/415 nm respectively. Moreover, the fluorescence peaks are far away from the other PAHs. There appear continuous peaks in the acenaphthene solution where the emission wavelength is in the range of 300~485 nm, and the corresponding excitation wavelength is 255~360 nm. The PAHs prediction error of N-PLS method for different water quality is small, where the RMS error of phenanthrene and fluorene are less than 0.4 μg·L-1, the relative error is less than 6%, and the RMS error of acenaphthene and fluoranthene are less than 1.0 μg·L-1, their relative error are less than 9%. The diffusion degree of PAHs is determined by the simulation and analysis of the diffusion tendency of four different kinds of PAHs in river, where the diffusion rate of fluorene and phenanthrene is about 51 mg·L-1, and acenaphthene and fluoranthene is 21 mg·L-1. Their diffusion rate is linear in a certain range and there is a linear relationship between PAHs and its concentration in accordance with Lambert-Beer law. The iteration times with the highest RMS error accuracy are obtained through the RMS error analysis of N-PLS method with different iteration times. The fit and correlation coefficient of the N-PLS method for PAHs prediction with different main factor numbers are compared The results showed that when the number of main factors is 3, the fitness could be up to 96.5, and the effect of N-PLS prediction model is optimal. Overall, the proposed method has higher detection accuracy, better recovery rate and stronger robustness compared with other detection methods.