[Objective] We aimed to solve the problem of monitoring data noise and boundary ambiguity in comprehensive evaluation of agricultural water quality, in order to establish a comprehensive evaluation model with good disturbance resistance and grade division. [Methods] A data-driven fuzzy support vector evaluation method was proposed to determine index weight of projection pursuit index and the parameters of fuzzy membership. Improved genetic algorithm was adapted to optimize the projection pursuit function and obtain the relatively objective index weigh. Then the parameters of fuzzy membership were optimized with data, and a comprehensive evaluation model of fuzzy support vector machine was constructed to reduce the influence of monitoring noise on the generalization ability of the evaluation model. In addition, considering the low resolution of the general discrete evaluation grade, the concept of regional division reliability was proposed to explain the reliability of the regional division grade of the sample, to further explain the comprehensive evaluation results. [Results] The model evaluation results were consistent with the results from experts and traditional evaluation. The model maintained more than 85% consistent rate with the monitoring data with 10%～30% random noise, and the reliability of regional division of samples was greater than the critical value, indicating the reliability and robustness of the method. The results from the constructed model were better than the fuzzy comprehensive evaluation and grey clustering method. [Conclusion] The method proposed by the present study is feasible and robust, and it can provide a reference for real-time evaluation of agricultural water quality under the condition of subsequent noise.