Water resource management requires forward-looking guidance, yet most water quality studies focus exclusively on retrospective analysis without providing probabilistic forecasts essential for proactive planning. This study presents a comprehensive medium-term forecasting framework combining Prophet time series algorithms with gradient boosting ensembles to project water quality trajectories through 2030 for Brazil's Paraíba do Sul Basin, a critical multi-jurisdictional watershed serving 14 million inhabitants. Using 14 years of monitoring data (2012–2025, n=3,559 measurements, 192 stations), we developed and validated multiple forecasting approaches, quantified prediction intervals, and evaluated alternative management scenarios. The gradient boosting ensemble achieved exceptional short-term prediction performance with test R² = 0.980, MAE = 1.17, and RMSE = 1.66, outperforming individual models by combining complementary algorithmic strengths through meta-learning. Among individual algorithms, LightGBM demonstrated superior accuracy (R² = 0.984, MAE = 0.98), followed by XGBoost (R² = 0.979, MAE = 1.11) and the base gradient boosting implementation (R² = 0.974, MAE = 1.23). Feature importance analysis revealed dissolved oxygen (32.3%), turbidity (25.7%), and biochemical oxygen demand (23.6%) as dominant predictors, collectively explaining 81.6% of forecast variance. Prophet algorithm provided complementary medium-term forecasting capabilities with uncertainty quantification, projecting continued water quality decline from current mean IQA of 74.4 to 69.8 by 2030 under business-as-usual scenarios, representing a 4.6-unit decrease driven by declining trends identified across all three basin states (São Paulo: −1.01 yr⁻¹, p=0.009; Rio de Janeiro: −0.67 yr⁻¹, p=0.10; Minas Gerais: −0.58 yr⁻¹, p=0.17). This trajectory appears incompatible with regulatory goals and basin management targets. The integrated modeling framework demonstrates that machine learning approaches provide valuable forecasting capabilities for large-scale water quality systems, with gradient boosting ensembles excelling at point prediction accuracy and Prophet offering interpretable trend decomposition with probabilistic projections suitable for risk-informed planning.