Blue-green algae are ancient organisms capable of photosynthesis with strong vitality. However, the toxins they produce pose a threat to human health and water safety. Global warming and water pollution have led to frequent outbreaks of algal blooms worldwide. Therefore, accurate prediction of blue-green algae concentration and early warning is crucial for effective algae management. However, current prediction models often lack accuracy and robustness, limiting their effectiveness in real-world applications. To enhance cyanobacterial bloom prediction accuracy, this paper proposes a time series prediction model for cyanobacterial concentration. The main structure includes wavelet decomposition, a many-objective optimization algorithm, and Gated Recurrent Unit neural network. The main process involves three steps: initially, the cyanobacteria concentration data and environmental variables undergo discrete wavelet decomposition to extract low-frequency trends and high-frequency features. Then, the decomposed data are fed into a two-layer GRU network, fine-tuned by the many-objective optimization algorithm for prediction. Lastly, the predicted outputs are combined to derive the final concentration prediction. To validate the model, we selected three metrics: RMSE (root mean square error), NSE (Nash-Sutcliffe efficiency coefficient), and CORR (correlation coefficient), to evaluate the prediction results. Meanwhile, we conducted tests in three regions: Morgan, Murray Bridge, and Tailem Bend, with RMSE values of 2721, 672, and 470 respectively; NSE values of 0.962, 0.979, and 0.969 respectively; and CORR values of 0.984, 0.990, and 0.986 respectively. The results demonstrate the generalizability of the model.