Models for managing water and chemical reagent injection based on machine learning and their impact on the preservation of hydrological balances and surface water quality

Abstract

The study is devoted to the development and evaluation of hybrid machine-learning-based models for managing water and chemical reagent injection aimed at reducing environmental risks and preserving hydrological balances during industrial exploitation of underground horizons. The goal is to quantitatively demonstrate how a predictive–optimization control loop can simultaneously enhance the technological stability of the process and improve surface water quality by minimizing diffuse and accidental discharges. Materials and methods include a synthetically realistic dataset of more than 2.5 million records over a five-year period with an hourly time step, combining injection telemetry, downhole equipment parameters, and hydrochemical monitoring; the Python/Pandas/NumPy/Scikit-learn/TensorFlow stack was used to build an ensemble regression model (XGBoost) for 24–72-hour forecasting of BOD5, COD, chlorides, sulfates, and suspended solids; an LSTM anomaly classifier for early leak detection; and an optimization addon based on a genetic algorithm for selecting flow rate, pressure, and composition considering the objective function “efficiency–environmental damage.” Verification was carried out using hold-out and k-fold (k=5) methods. Results show technological benefits: reduction of average consumption of reagent A by 19.05% and reagent B by 11.30%, decrease in injection pressure variation coefficient by 72.32%, reduction of unplanned shutdowns by 78.57%, and decrease of unproductive injection portion by 71.41%; environmental effects include reductions in average annual chloride concentrations from 48.17 to 21.33 mg/L, sulfates and suspended solids by 33–43%, and return of petroleum products and BOD5 to background levels at the control section. Correlation analysis records the disruption of strong relationships between process instability and pollution (r up to 0.83 under conventional control → r less than 0.2), and groundwater level forecast accuracy reaches an average relative error of 1.66%, allowing for proactive water balance maintenance. The discussion emphasizes the transition from reactive to proactive management, the synergistic effect of reduced variability and risk, and the prospects of scaling through integration with digital twins and reinforcement learning methods.