Spectral confocal technology realizes the optical non-contact precision measurement of micro displacement based on the principle of dispersion and confocal. The measurement accuracy of the spectral confocal technology is limited by the extraction accuracy of the spectral peak wavelength. The General Regression Neural Network (GRNN) is proposed for the spectral characterization of the spectral confocal system and the precise positioning of peak wavelength. The GRNN model is a feedforward network model with simple structure, concise training, and fast convergence speed. The proposed spectral characterization algorithm is verified on established the spectral confocal experimental system. The precision displacement table moves the mirror to the zero working position of the dispersion probe, and the spectrometer detects the dispersion spectrum focused and reflected from the mirror surface. Denoising and intensity normalization is performed on the collected original spectral data, and some spectral signals data in the spectral range near the peak are intercepted and input into the GRNN model as sample data. The input variable is the signal wavelength λ in the spectral data. The normalized intensity of the spectral wavelength is the output variable. The joint probability density function of the input and output variables of the sample is the verification condition. Finally, the GRNN model outputs the maximum probability value of normalized intensity corresponding to wavelength through Parzen nonparametric kernel regression. The GRNN model considers the weight of sample points near output variables in spectral characterization, and it can eliminate the influence of random noise of spectral signals, improve the spectral signal-to-noise ratio, reduce the characterization error of spectral signals, and improve the accuracy of peak wavelength extraction to achieve stable and reliable spectral confocal measurement. Then the wavelength of the spectral peak is extracted in different dispersion positions by the GRNN model. The corresponding relationship between the dispersion wavelength and the focus position is revised. Experimental results show that the GRNN model is better than the traditional algorithms. The spectral fitting curve SNR of the GRNN model is improved. The fit coefficient of the fifth-order dispersion focal shift is 0.999 9. The system resolution is about 2 μm. The measurement error RMSE is about 0.01 μm. The GRNN model suppresses the dispersion model fluctuation caused by wavelength extraction and improves the resolution and stability of system measurement.