• • CNN-BiLSTM-Attention achieved the best performance for water temperature prediction, with R² exceeding 0.95 and RMSE below 0.5°C, making it suitable for stable, low-complexity forecasting scenarios.
• • TCN-Transformer outperformed other models for pH, dissolved oxygen, and turbidity, achieving R² improvements of 8-12% over SVR and XGBoost, and reducing RMSE by up to 20% for turbidity, demonstrating its capability in capturing nonlinear and long-range dependencies.
• • Combining IQR outlier removal with multiple imputation strategies reduced prediction error (MAE) by 15-25% across all models, highlighting the critical role of data quality enhancement in water quality forecasting.
• • Bayesian optimization consistently improved model performance, with average R² gains of 0.03-0.05 compared to default hyperparameters, ensuring robust and reliable predictions for operational deployment.