A Real-Time Multi-Task Deep Learning Framework For Smart Water Quality Prediction

Water quality prediction is essential for effective water management, public health, and environmental protection, especially with the increasing pollution and climate change issues. This paper presents RealHAXNet, an optimized multi-task learning model that delivers real-time predictions of five water quality indicators simultaneosly, including both sensor-measurable and indirectly measurable parameters. RealHAXNet builds on the complex and high-performing HAXNet archiecture, optimizing it with a fusion of pruning and quanitzation methods to reduce model size by 65% and reduce real-time inference latency from 81 ms to less than 6 ms, while also either maintaining or even improving prediction performance. Evaluated on a large dataset from 58 water reservoirs in Spain, RealHAXNet achieves an average RMSE of 0.0186 and R² of 0.9877, outperforming the original model by up to 63%. The framework also simulates real-time sensor data streams, performs real-time SAITS imputation, and also incorporates online incremental learning and drift detection to improve model adaptability over time. These results demonstrate RealHAXNet’s potential for scalable, efficient, and adaptive water quality monitoring on resource-constrained devices.