Machine learning–based modelling of effluent quality of Wupa municipal wastewater treatment plant, Abuja, Nigeria

Accurate prediction of effluent quality remains a significant challenge in wastewater treatment due to fluctuating influent loads, sensor noise, and non-linear system behaviour. This work evaluates three machine-learning models, Decision Tree, Random Forest, and XGBoost, using a five-year dataset (1825days) comprising daily influent and effluent measurements of pH, TDS, TSS, BOD, COD, and microbial indicators. A comprehensive preprocessing framework, including noise reduction and smoothing, reduced parameter variability by 28–51% and revealed clearer long-term patterns essential for modelling. Comparative analysis showed that Random Forest provided the most reliable generalisation, achieving test R² values of 0.414 (pH), 0.317 (TDS), 0.221 (BOD), 0.200 (COD), and 0.133 (TSS), outperforming XGBoost, which despite high training accuracy (e.g., R² = 0.93 for pH), suffered substantial overfitting with test R² dropping to as low as 0.046 for E. coli. Decision Tree performed the weakest overall, with consistently low predictive accuracy across all parameters. Effluent trend analysis indicated strong operational performance, including high removal efficiencies for BOD (92.7%), COD (89.1%), TSS (91.1%), and microbial indicators (~99%), while TDS removal remained limited (29.4%). All models struggled to predict microbial parameters due to their episodic variability and weak correlation with physicochemical inputs. Overall, Random Forest emerged as the most robust algorithm for effluent forecasting. At the same time, results highlight the need for advanced feature engineering, additional process-level data, and specialised modelling approaches to improve microbial prediction. These findings demonstrate the potential of machine-learning tools to enhance real-time monitoring, early anomaly detection, and decision support in wastewater treatment plants.