Predicting the erratic interannual fluctuation of harmful cyanobacterial bloom in response to environmental changes is critical for ecosystem-based management. The predictability would largely depend on to what degree the underlying process of bloom dynamic is forced by random noises or deterministically structured factors. Here, we analyzed a long-term cyanobacteria, Microcystis spp., time series data with the environmental variables over 30 years in Lake Kasumigaura, Japan. We found that (1) the Microcystis abundance signals as low-dimensional, nonlinear, and chaotic, which can be distinguished from random noises at an interannual scale, (2) the out-of-sample forecasting of time series based on the nonlinear method out-performs the linear method (3) the dynamical causality between bloom dynamic and ecosystem variables can be identified through state space reconstruction. Our findings reconstruct the causal relationships between cyanobacterial bloom and its determinants, and represent a precautionary approach for annual bloom forecasting under environmental changes.