AI prediction of ammonium levels in rivers using machine learning

A developed machine learning framework for predicting ammonium (NH₄⁺) levels in River Lee, London, is presented in this paper. We use state-of-the-art algorithms such as the Temporal Fusion Transformer (TFT), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) on a large dataset that includes temperature, turbidity, chlorophyll, dissolved oxygen, conductivity, and pH. By clarifying the intricate relationships between environmental variables and ammonium levels, these models greatly improve forecast accuracy. Using the TFT model for multi-horizon forecasting is one of our research's unique features. High accuracy and interpretability in hydrological predictions are made possible by this model's skilful integration of convolutional elements with an attention mechanism. It solves a crucial problem in environmental modelling by skilfully managing short-term variations while being resilient over longer periods of time. Adaptability and resilience are combined in our dual-scale method, which works well for both short- and long-term environmental projections. In particular, XGBoost performs exceptionally well in monthly forecasts up to 12 months with a noticeably low RMSE, while the RF model exhibits exceptional long-term forecasting capabilities, attaining an R² of 0.97 and an RMSE of 0.18 over 1095 days. TFT performs best in short-term projections, but data granularity limits its ability to perform well in longer-term situations. These revelations highlight how urgently proactive water management techniques are needed to reduce hazards like hypoxia and possible ecological effects. In the end, our research offers resource managers vital assistance in tackling issues pertaining to ammonium toxicity and ecological health.