Research on water quality monitoring and abnormal early warning from the dual-driven perspective of knowledge and data

Water resources are the foundation for human survival and development, yet intensified human activities have increased the risks of water quality deterioration. Existing water quality early warning studies predominantly emphasize prediction performance while often under-addressing anomaly determination, which contributes to frequent false alarms in operational monitoring. From a dual-driven perspective of knowledge and data, this study develops an integrated framework for water quality monitoring and abnormal early warning. On the knowledge-driven side, abnormal water quality observations are systematically categorized into pseudo anomalies induced by accidental disturbances and true anomalies characterized by sustained structural deviations such as violent fluctuations or upward/downward leaps. Accordingly, reproducible identification strategies are designed, including rolling-statistics-based detection for isolated false data points, rolling-variance screening for variance-shrinkage sequences, and difference-threshold rules for rapid-change segments, enabling the exclusion of pseudo anomalies from model learning and improving alert credibility. On the data-driven side, a Stacking ensemble learning model is constructed by integrating multiple base learners to capture nonlinear temporal patterns and enhance multi-step prediction performance. The proposed framework is validated using high-frequency water quality, water level, and rainfall data from three monitoring stations in the Phase II of the Eastern Route of China's South-to-North Water Diversion Project. Results show that the ensemble model demonstrates good tracking ability, while the knowledge-data coupling mechanism effectively distinguishes pseudo anomalies from true anomalies, thereby reducing false alarms and improving the practical feasibility of water quality monitoring and early warning.