Real-Time COD Prediction in Full-Scale Wastewater Treatment Using Ensemble Learning Models and Systematic Feature Selection

Accurate and timely prediction of Chemical Oxygen Demand (COD) is essential for efficient wastewater treatment plant (WWTP) operation; however, conventional laboratory-based COD measurements are time-consuming and unsuitable for real-time process monitoring. This study develops and compares three ensemble machine learning models—Random Forest (RF), XGBoost, and LightGBM—for COD prediction using six years of operational data collected from the full-scale Melbourne Eastern Treatment Plant. A multi-stage feature selection framework combining Average Rank aggregation, Pearson correlation analysis, and Sequential Backward Elimination was employed to reduce the original ۱۹-variable input space to an optimal subset of five predictors: Total Nitrogen (TN), Biological Oxygen Demand (BOD), year, mean air temperature, and minimum air temperature. Model hyperparameters were optimized using Randomized Search with Time Series Split cross-validation, and predictive performance was evaluated using a chronological train–test split. The results demonstrated that feature selection improved model generalization while reducing model complexity. Among the evaluated algorithms, LightGBM achieved the highest predictive performance on the test dataset (R² = ۰.۷۱۶, RMSE = ۷۶.۳۰ mg/L, and MAE = ۵۶.۵۹ mg/L), outperforming XGBoost (R² = ۰.۶۸۲) and RF (R² = ۰.۶۵۳). TN and BOD consistently emerged as the most influential predictors across all models, highlighting their strong relationship with COD dynamics in wastewater treatment processes. Overall, the findings indicate that LightGBM combined with a systematic feature selection strategy provides an efficient and practically deployable framework for near-real-time COD prediction, offering a promising alternative to conventional laboratory analyses for WWTP monitoring and control.

کلیدواژه ها:

Chemical oxygen demand ، Wastewater treatment plant ، Ensemble machine learning ، Feature selection ، LightGBM