Leakage-Controlled Comparative Prediction of Microplastics in Sewage Treatment Plant Effluents: Season- and Morphology-Aware Modelling at Erandwane STP, Pune, India

Microplastics (MPs) discharged in sewage treatment plant (STP) effluents vary with influent concentration, particle morphology and temporal conditions, but predictive studies can be weakened by data leakage, limited benchmarking and mechanistic overinterpretation. This study evaluates within-site prediction of final-effluent MP concentration at the Erandwane STP, Pune, India, using 792 contemporaneously paired influent–effluent observations from 396 dates between 1 August 2024 and 31 August 2025. Predictors were total influent MPs, influent fibres, fragments, films and beads, categorical hydrological season and sampling session. Morning and afternoon records from the same date were kept in the same modelling subset to prevent same-day leakage, and five-fold date-grouped cross-validation compared multivariable linear regression, support vector regression, random forest, XGBoost and CatBoost. On the independent grouped test set, CatBoost achieved a coefficient of determination (R2) of 0.498, a root mean square error (RMSE) of 35.52 MPs L−1 and a mean absolute error (MAE) of 28.58 MPs L−1; however, its mean grouped cross-validation R2 (0.433) did not exceed multivariable linear regression (0.473). The principal finding is therefore moderate site-level predictive information in the available multivariable structure rather than superiority of a specific algorithm. Season was the leading CatBoost importance variable but is interpreted only as a temporal proxy because direct rainfall, flow, hydraulic retention time, clarifier loading and total suspended solids were not continuously available. Quantitative recovery, duplicate-precision and detection-limit data were also unavailable, and the same validation subset was used for tuning, calibration and conformal-residual estimation. Consequently, the uncertainty results and model interpretation should be considered exploratory. The framework may support prioritisation of confirmatory monitoring at this site, but it cannot replace laboratory measurement or establish transferable treatment mechanisms without prospective, multi-site validation and stronger analytical and operational covariates.