A selective model combination method is proposed in this paper to improve the precision of water quality analysis with three dimensional fluorescence spectra. A correlation coefficient criterion was designed to select effective excitation wavelengths for sub-models building, based on which the ridge regression method was adopted to combine the selected sub-models to get the stacked model. Thirty two samples from surface water and urban wastewater were used as research objects with total organic carbon (TOC) index from 3.41 to 125.35 mg·L-1, and chemical oxygen demand (COD) index from 22.80 to 330.60 mg·L-1, and 10 excitation wavelengths in the range of 220-400 nm were adopted to generate three dimensional fluorescence spectra. Following the proposed correlation coefficient criterion, the excitation wavelengths of 260, 280 and 400 nm, and the excitation wavelengths of 220, 280 and 400 nm were selected respectively for TOC analysis and COD analysis, based on which two stacked models were built by using partial least square regression method for sub-models building and ridge regression method for sub-models combination. The experimental results show that, compared with the sub-models with the best prediction precision, the root mean square errors of prediction (RMSEP) of the stacked models decreased by 15.4% for TOC analysis, and 17.5% for COD analysis; and compared with the models without sub-models selection, the RMSEP of the stacked models decreased by 6.1% for TOC analysis and 10.9% for COD analysis