[Objective] It is one of the key basic issue in the prevention and control of water environment by exploring effective prediction methods about cyanobacteria bloom in lakes and reservoirs.[Methods] Combined with the class random characteristic showed in the chaotic evolution of cyanobacteria bloom, this paper proposed a dynamic clustering algorithm based on the optimization of validity functions to achieve the optimal cluster number of cyanobacteria bloom and small-scale neighborhood optimal prediction. First of all, monitoring data were classified objectively by the proposed dynamic clustering algorithm to reduce effectively the search space and to improve the prediction accuracy. Then the optimal number of neighbors for all kinds was obtained using the particle swarm optimization algorithm, which was used to determine the number of participating in the local regressive algorithm. Finally, a dynamic regressive prediction model was established.[Results] The model established using the concentration data of chlorophyll a at the Jinshu monitoring site of Taihu Lake in 2011 was used to model and predict short-term variation of it in 2012. The predicted value of the model was consistent with the actual trend and the relative error was 12.02%, and was smaller than the ones predicted by other models, such as linear regression algorithm based on traditional clustering, BP neural network, and phase space reconstruction algorithm, whose relative errors were 15.21%, 19.51% and 38.42%.[Conclusion] Numerical results showed that the prediction accuracy of this method was relatively high, hence the feasibility and effectiveness of the optimization prediction method proposed were proved.