Advancing river water quality prediction: a comparative assessment of deep learning models for dissolved oxygen forecasting

Abstract

Accurate forecasting of dissolved oxygen (DO) is crucial for monitoring river water quality and protecting aquatic ecosystems. This study compares the performance of four deep learning models – Temporal Fusion Transformer (TFT), Informer, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) – for forecasting DO concentrations in the River Lee (London, UK) across 7- and 30-day time frames. A multivariate time-series dataset was employed, with temperature, turbidity, pH, conductivity, chlorophyll, and river flow as predictors. Model skills were evaluated using RMSE, MAE, R2, and SMAPE. Over the 7-day period, TFT had the lowest RMSE (0.06) and SMAPE (8.86%), while LSTM had the greatest R2 (0.77). TFT outperformed Informer, LSTM, and GRU at the 30-day horizon, with R2 = 0.79 and SMAPE of 8.23%, despite significant accuracy losses. According to the variable contribution study, temperature and river flow were the most significant factors, particularly for short-term projections. Overall, the results show that transformer-based structures, particularly TFT, can successfully represent nonlinear temporal dependencies and multivariate interactions, making them ideal for multi-horizon DO forecasting in river systems. These models have the ability to supplement normal monitoring by offering short-term predictions about probable oxygen conditions.