Predicting water quality is crucial for health and environmental management. This study developed a machine learning-based framework for predicting Total Phosphorus (TP) and Ammonia Nitrogen (NH₃-N) concentrations in the Qinhuai River (QR) basin. The framework integrates three machine learning models: Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM), and multiple data-driven scenarios to forecast water quality for 24, 48, and 72 h. Three data-driven input scenarios were designed: water quality, hydrology, and rainfall parameters. Results show that the LSTM model performed best. For example, it achieved high accuracy (Nash-Sutcliffe Efficiency, or NSE = 0.84) in predicting NH₃-N one day ahead, and maintained strong performance (NSE = 0.72) even for 72-h TP forecasts. Performance declined with more input variables, indicating feature redundancy harms generalization. NH₃-N predictions were more accurate than TP and more sensitive to lead time. Spatially, the upper reach showed stable performance due to natural dominance; the middle reach captured a TP decline from 0.12 to 0.10 mg/L; the lower reach was most challenging due to tidal and pumping effects; and the part affected by water regulation showed high stability and optimal model performance. The framework performed well in downstream Yangtze River stations, highlighting its potential for broader water quality monitoring and watershed management.