Advanced air pollutant forecasting: Integrating wastewater effluent data with ensemble deep learning (LSTM-XGBoost/RF)

Abstract This study introduces a novel framework for forecasting ambient air pollutant concentrations (hydrogen sulfide, H₂S, and nitrogen dioxide, NO₂) using routinely monitored wastewater effluent parameters, offering a cost-effective alternative to unstable and expensive sensor-based air quality monitoring. We investigated statistical relationships between H₂S and chemical/biochemical oxygen demand (COD/BOD₅), and between NO₂ and nitrogen species (NO₃⁻/NH₃) at two wastewater treatment plants in Jordan from 2016 to 2020. The analysis revealed strong, though year- and site-specific, positive correlations (Pearson r up to 0.99 for H₂S–COD and up to 1.0 for NO₂–NO₃⁻), highlighting the potential of wastewater parameters as predictive proxies. To translate these relationships into robust forecasting tools, we developed and compared a Multiple Linear Regression (MLR) baseline with advanced ensemble models, including Random Forest (RF) and a hybrid Long Short-Term Memory–XGBoost (LSTM–XGBoost) architecture. The LSTM–XGBoost model demonstrated superior performance by effectively capturing both long-term temporal dependencies and complex non-linear interactions, accurately predicting seasonal trends such as summer H₂S peaks (0.9–19.5 ppb) and elevated NO₂ during wet periods. On a held-out test set, it achieved high predictive accuracy (R² > 0.90 for NO₂, R² > 0.85 for H₂S) with low error (RMSE < 1.1 ppb), outperforming both standalone models and traditional regression. This integrated approach establishes a stable, data-driven framework for proactive air quality management at wastewater facilities, significantly reducing dependence on frequent sensor calibration and maintenance while enabling timely pollution mitigation strategies.