Long-Term Forecasting of Dissolved Oxygen in Rivers Using Machine Learning Models

Predicting dissolved oxygen (DO) levels in river ecosystems—particularly in River Lee, London—is crucial to maintaining aquatic life and water quality. Using daily data, this work presents machine learning models that can forecast DO levels over a variety of periods, from short (7 days) to long (365 days). We enhanced the capacity of long-term DO forecasting by utilizing models such as Temporal Fusion Transformer (TFT), Informer, Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). With an RMSE of 0.04 and an R2 of 0.09 at the 365-day horizon, the Informer model performs well in managing long-term dependencies. On the other hand, although the TFT model consistently performs well throughout a range of time periods, the LSTM and GRU models' accuracy decreases for forecasts longer than 90 days. Furthermore, DO levels are greatly influenced by environmental factors such as pH, chlorophyll, turbidity, temperature, conductivity, and river velocity. Environmental organisations can develop proactive water management plans and prevent problems like river hypoxia thanks to the enhanced performance of models like the Informer and TFT. These results highlight how cutting-edge machine learning methods can help ensure the long-term viability of river ecosystems.

 

Keywords: River streamflow; LSTM; GRU; TFT; Informer; Water Quality