<p>The paper industry is a water-intensive and highly polluting sector, and its wastewater treatment process poses major challenges for accurate effluent quality prediction. To address the strong nonlinearity, non-stationarity, noise contamination, and multi-scale fluctuations in effluent chemical oxygen demand (COD) series caused by feedstock switching and operational disturbances, this study proposes a two-stage optimized decomposition–integration forecasting framework. In the first stage, a Genetic Algorithm-optimized Variational Mode Decomposition (GA•VMD) method adaptively extracts informative multi-scale components from the raw COD series, reducing mode mixing and improving feature representation. In the second stage, Bayesian-optimized Bidirectional Long Short-Term Memory (BO•Bi-LSTM) networks automatically tune hyperparameters for each decomposed component in a multivariate forecasting framework with auxiliary process variables. A case study based on operational data from a paper-mill wastewater treatment plant in Zhaoqing, China, shows that the proposed framework achieves the best overall performance among all compared models, achieving a mean absolute error (MAE) of 0.3127&#xa0;mg·L^⁻1, root mean square error (RMSE) of 0.6134&#xa0;mg·L^⁻1, mean absolute percentage error (MAPE) of 0.8620%, and a coefficient of determination (R²) of 0.9431. Compared with the strong BO•Bi-LSTM baseline, the proposed framework reduces MAPE by 30.1% and RMSE by 32.8%, indicating clear advantages in average-error reduction and suppression of large prediction deviations. Four-fold walk-forward validation and Diebold–Mariano statistical tests provide supplementary evidence of its robustness under the present case-study conditions. The proposed framework provides an effective soft-sensor tool for supporting proactive chemical dosing, operational adjustment, and smart wastewater treatment management.</p>