Bridging Data Gaps and Forecasting Challenges in Water Quality Monitoring using Deep Learning

Reliable water quality monitoring is critical for protecting public health and enabling sustainable environmental management. However, high rates of missing data often exceed 40% due to sensor failures and operational disruptions severely undermine the utility of environmental monitoring systems. To address this challenge, this thesis presents a three-stage deep learning framework designed to enhance the accuracy and resilience of water quality forecasting. First, a Bidirectional Long Short-Term Memory model with Attention and Autoencoder Modules (BiLSTM-AAM) is introduced to impute long, irregular gaps in multivariate time series. Applied to real-world data from the St. Lucie River Basin, the model achieved a Nash-Sutcliffe Efficiency (NSE) of up to 0.9942 while handling 41.8% missingness. Second, a Modular BiLSTM architecture is developed to perform high-fidelity multi-stream prediction using gap-filled data. Evaluated on ammonium (NH₄⁺) forecasting from a clay-perlite with sand (CPS) filtration cell, it achieved an R² of 0.9857 and reduced MAE by 73.7% compared to baseline models. Finally, the Reinforcement Learning-enhanced Modular BiLSTM (RL-MoBiL) framework is proposed to dynamically correct prediction errors, particularly for volatile parameters like nitrate (NO₃⁻). RLMoBiL achieved an R² exceeding 0.999 on CPS-derived data. Collectively, these contributions form a robust, adaptive solution to the persistent issue of missing data in water quality monitoring and establish a foundation for more intelligent, resilient, and accurate environmental forecasting systems.