Machine Learning Pipeline for Wastewater Effluent Quality Forecasting: Gradient Boosting-Based BOD5 Prediction with Feature Importance and Seasonal Pattern Analysis
TL;DR
A machine-learning-based predictive framework integrating temporal feature engineering, ensemble gradient-boosting models, and explainable artificial intelligence for multimonth BOD5 prediction is proposed, enabling reliable predictions up to 12 months ahead.
Abstract
Realtime BOD5 forecasting is essential for optimizing wastewater treatment operations and ensuring regulatory compliance. This study proposes a machine-learning-based predictive framework integrating temporal feature engineering, ensemble gradient-boosting models, and explainable artificial intelligence for multimonth BOD5 prediction. Monthly wastewater quality data from 2008 to 2023 (192 months), including BOD5, total suspended solids, total phosphorus, and flow rate, were used to construct 48 temporal features based on lagged variables, rolling statistics, and physicochemical ratios. Four ensemble models─Gradient Boosting, Random Forest, XGBoost, and LightGBM─were trained on 2008–2020 data and evaluated on independent data from 2021 to 2023. LightGBM achieved the highest accuracy (R2 = 0.9894, RMSE = 0.223 mg·L–1, MAPE = 2.26%), enabling reliable predictions up to 12 months ahead. Feature importance analysis identified six month lagged BOD5 and the BOD5-to-flow ratio as dominant predictors. The framework supports proactive effluent quality management and early warning decision support.