Water quality prediction method based on a combined machine learning model: A case study of the Daling River Basin

Abstract

Total nitrogen (TN) is a key factor limiting river water quality. Identifying its spatial-temporal variation and influencing factors is essential for predicting pollution trends and mitigating water quality risks. However, challenges such as insufficient real-time monitoring, limited spatiotemporal modeling, and complex feature selection for multi-source data remain. This study proposes a machine learning method combining spatiotemporal weighted interpolation, relevant feature selection, and time-series decomposition to analyze TN concentration variations in the Daling River Basin (2023–2025) using data from seven monitoring stations. The Enhanced Long Short-Term Memory with Back Propagation Network (ELSTM-EBP) model is developed to assess TN spatial-temporal fluctuations and driving factors. Results show that: (1) TN concentrations are highest in January and February, with a “U”-shaped fluctuation over the year, and lowest in August; (2) Significant spatial heterogeneity is observed at the seven monitoring points; (3) Water temperature negatively correlates with TN, while dissolved oxygen positively correlates; other factors such as the permanganate index and turbidity also significantly influence TN levels; (4) The ELSTM-EBP model outperforms other models (ELSTM-EBP > QLSTM > LSTM > GRU-QIMAS > EQINN > BP) in TN prediction accuracy and generalization ability; (5) Multi-step prediction accuracy decreases slightly with step length, but remains within −0.4 to 0.4 mg/L for up to 7 steps, indicating robust performance for short-term predictions.