The water quality parameters in the water environment are non-linear, random and dependent. The prediction accuracy and robustness of traditional water quality prediction models are generally not high. In order to optimize and improve the prediction accuracy of water quality prediction models, this paper proposes a multi-source K-means clustering combination GA-BP neural network is used to study the dynamic classification and prediction of water quality. Integrated learning uses multiple learning algorithms to obtain better prediction performance than traditional single learning algorithms. First, the water quality elements are classified according to similarity through multi-source K-means clustering, and then the weight of each element is calculated through the classification results. The GA-BP neural network is used to predict the changes of various elements of water quality. The application of the example of 36 feet Lake in Pingtan, China shows that the method is effective and feasible, and the accuracy of prediction is obviously improved which is helpful for analyzing the water quality of 36 feet Lake.