Abstract: To address the limitations of traditional models in terms of accuracy and timeliness, this study established an intelligent water quality prediction model for the Danjiangkou Reservoir based on the EEMD-LSTM algorithm. The model enhances data quality and optimizes feature selection through a three-stage progressive preprocessing method and a dual-dimensional feature analysis. Subsequently, Ensemble Empirical Mode Decomposition (EEMD) is employed to suppress mode mixing and extract multi-scale temporal features. Finally, the model integrates the capability of Long Short-Term Memory (LSTM) networks to capture long-term dependencies, thereby achieving accurate prediction. Experimental results demonstrate that for total phosphorus predictions across five monitoring stations, the model achieves a short-term (1–3 days) coefficient of determination (R²) exceeding 0.8 and a medium-term (7 days) R² above 0.65. The findings indicate that the EEMD-LSTM model effectively mitigates the prediction lag inherent in standalone LSTM models, significantly enhancing the accuracy and stability of water quality prediction. This model can provide robust technical support for real-time water quality forecasting and risk management in the Danjiangkou Reservoir.