Accurate prediction of water quality parameters in a water treatment system remains challenging due to the intertwined physical, chemical, and microbiological nature of diverse data streams. We propose a modular bidirectional LSTM architecture that integrates data from water quality sensors, microbial indicators, and weather measurements to handle the prediction complexity, when preserving domain-specific temporal patterns. Evaluated on a specific filtration media in such a water treatment system in Martin County, Florida, our approach achieves significant prediction accuracy for ammonium (NH 4 +) prediction in the effluent over traditional methods, with R 2 = 0.9912 and RMSE = 0.0058. This represents an 85% reduction in mean absolute error compared to standard bi-directional LSTM approaches (from 0.0281 to 0.0041) with no modularized input streams. The architecture’s effectiveness stems from its ability to maintain distinct temporal patterns while enabling integrated prediction, demonstrating strength during periods of parameter volatility from dry to wet season. Results suggest that modular processing of heterogeneous environmental data streams provides a robust foundation for water quality management.
Gudla et al. (2026) studied this question.