In order to better deal with the uncertain information in the process of water quality evaluation, the theory of set pair analysis is used. Considering the water quality classification level refinement, we divide the diversity factor in set pair analysis theory into good-difference and bad-difference and divide the reverse factor into good-reverse and bad-reverse. The weight of the index is calculated by the combining weighting method and corrected by considering the influence of stochastic observation error. The difference coefficient and correlation are used to optimize the index system. This model is applied in the evaluation of ground water functional zone in Yixing. The results are compared with those obtained using the artificial neural network, gray theory, and projection pursuit methods. When the evaluation results of the four methods are different in some sections, we find that the results using this model are in the best agreement with the water quality monitoring data.