A hybrid framework for water-quality prediction using WOA-optimized EEMD and CNN-LSTM

Accurate prediction of dissolved oxygen is important for water-quality management, yet monitoring series are usually noisy, nonlinear, and non-stationary, which limits conventional forecasting models. This study proposes a WOA-EEMD-CNN-LSTM framework in which EEMD is used to decompose the original series into intrinsic mode functions and a residual component for noise reduction and multi-scale representation, WOA is used to optimize the key EEMD parameters and improve decomposition quality, and CNN-LSTM is used to capture local features and long-term temporal dependencies for component-wise prediction before final aggregation. The method was evaluated on a 4 h monitoring series from the Xiaojiang estuary and a public Refuge dataset. In the locked seed-123 evaluation, the WOA-EEMD-CNN-LSTM model achieved an RMSE of 0.32867 and an $R^2$ of 0.95070 on Xiaojiang, and an RMSE of 0.28874 and an $R^2$ of 0.93429 on Refuge. The resulting workflow provides entropy-guided parameter adaptation, causal multi-scale representation, component-wise learning, timestamp-aligned aggregation, and an auditable leakage-controlled evaluation process. These properties make the framework a practical and reproducible option for dissolved-oxygen forecasting in noisy water-quality monitoring series and provide a clear basis for later multi-seed and multi-site validation.