River water quality forecasting: a novel LSTM-Transformer approach enhanced by multi-source data

Water quality prediction holds crucial importance as a fundamental technical support for efficient water resource management and strong ecological protection. In this study, aiming to meet the pressing requirement for eutrophication prevention and control in the water body of the Changzhou section of the Beijing-Hangzhou Canal, a prediction model for total phosphorus (TP) and total nitrogen (TN) concentrations, driven by deep learning, was constructed. A comprehensive multivariate dataset was formed by combining automated water quality monitoring data within the basin, remotely sensed interpretations of land types, and meteorological factors. To improve data quality, the wavelet denoising (WD) technique was applied to eliminate data noise. Based on this, a hybrid WD-LSTM-Transformer model was proposed. Initially, this model utilizes the LSTM network to capture the time-series dynamic characteristics of TP and TN and extract their cyclic change patterns. Subsequently, it analyzes the complex nonlinear relationships among water quality parameters, meteorological conditions, and land use patterns via the multi-head attention mechanism of the Transformer. The SHAP method was used to identify significant variables and infer their influence on the fluctuations of TP and TN, thus enhancing the interpretability of the model. Experimental results indicated that the model could predict TP and TN with R² exceeding 0.9. Compared with four traditional models, it significantly improved the prediction accuracy. By evaluating the significance of features while maintaining a high prediction performance, this model provides a scientific basis for the analysis of pollution sources within the watershed.