Using predictive analytics to prevent unauthorized wastewater discharges and optimize the operation of treatment facilities at oil and gas sites

Abstract

The aim of the study is to demonstrate how predictive analytics based on machine learning methods can prevent unauthorized wastewater discharges and optimize the operation of treatment facilities at oil and gas sites by shifting management from a reactive to a proactive paradigm. The relevance is due to the high variability of wastewater composition and the growing risks of shock pollutant inflows. The empirical base includes more than 85,000 records (2019–2023) with 28 technological parameters: flow rate, temperature, pH, suspended solids, petroleum products, COD, BOD₅, sulfides, phenols, ammonium nitrogen, as well as indirect indicators of main process unit modes and data on reagent dosing and aeration. Preprocessing covered outlier removal (Isolation Forest), missing value imputation (k-NN Imputer), and normalization (Min–Max); EDA, lag and moving average generation, and Pearson/Spearman correlation analysis were performed for feature selection and multicollinearity reduction. Random Forest, gradient boosting (LightGBM), and LSTM were compared; validation used time-series cross-validation with RMSE, MAE, and R² metrics for regression and Accuracy, Precision, Recall, and F1 for classification. The LightGBM model led in forecasting COD with a 3-hour horizon (R²=0.941; RMSE=41.53 mgO₂/L), showing comparable accuracy to LSTM but lower computational cost; feature importance analysis revealed contributions from indirect production parameters. Simulation of forecast integration into the coagulant dosing loop demonstrated a reduction in average dose from 85.00 to 68.43 mg/L and direct reagent savings of 19.50%, along with quality improvement: residual suspended solids decreased from 22.54 to 18.91 mg/L, and standard deviation fell from 14.88 to 8.12 mg/L, improving stability and reducing the risk of breakthrough. The anomaly classifier provides early warning 3–6 hours in advance with Accuracy 0.988, Precision 0.932, Recall 0.965, and F1 0.948, minimizing missed critical events and false alarms. The conclusions confirm that a two-level predictive loop (continuous regression forecast + anomaly detector) reduces resource consumption, the probability of exceeding permissible concentrations, and lays the foundation for integration with process control systems and digital twins, with a future perspective of expanding the factor field through meteorological data.