Machine Learning-Based Prediction of River Water Quality: A Streeter-Phelps-Grounded Simulation Framework and Comparative Model Evaluation

River water quality is governed by the interaction of hydrology, temperature-dependent biochemical kinetics, and episodic pollution loading, making both long-term trend characterization and short-term event forecasting central problems in civil and environmental engineering. This paper (i) reviews the classical Streeter-Phelps oxygen-sag framework and the Water Quality Index (WQI) concept that together underpin most operational water-quality assessment; (ii) reviews the rapidly growing literature applying machine learning (ML) to water quality and dissolved-oxygen (DO) prediction; and (iii) reports two complementary, fully reproducible simulation-based experiments. First, a 20-year daily simulation, grounded in the standard Benson-Krause DO-saturation-temperature relationship, quantifies a thermally-driven decline in DO saturation capacity of 0.072 mg/L per decade under a modest (+0.35 °C/decade) warming trend, with temperature and DO saturation correlated at r = -0.997. Second, an hourly-resolution, 3-year Streeter-Phelps pollution-event simulator, driven by episodic biochemical oxygen demand (BOD) loading pulses analogous in mathematical structure to the injection/decay processes used in companion geophysical forecasting studies, was used to train and evaluate Random Forest, Gradient Boosting, Multilayer Perceptron, and linear baseline models predicting WQI at 6-, 24-, and 72-hour horizons under a chronologically leakage-aware protocol. The best model (Gradient Boosting) achieved RMSE = 3.62 WQI points and R² = 0.441 at 6 hours, degrading to R² ≈ 0.10-0.13 by 24-72 hours, a steeper skill decay than reported for single-variable geophysical indices, attributed to the compounding of independent noise sources across the five WQI sub-indices. Feature-importance and ablation analysis show that recent WQI history dominates short-horizon predictability, while hydrological drivers (temperature, flow) alone are competitive at longer horizons. Because live access to real river-monitoring archives was not available in this environment, both experiments are explicitly disclosed as physically-grounded simulations rather than analyses of observational data, and the paper concludes with a discussion of the sim-to-real gap, operational relevance for water treatment and pollution-control decision-making, and directions for future work. Keywords: machine learning; neural networks; environmental engineering; Random Forest; Water Quality Index; Streeter-Phelps model; dissolved oxygen; pollution forecasting