The accurate prediction of total phosphorus (TP) is crucial for the early detection of water quality eutrophication. However, predicting TP concentrations among canal sites is challenging due to their complex spatiotemporal dependencies. To address this issue, this study proposes a GAT-Informer prediction method based on spatiotemporal correlations to predict TP concentrations in the Beijing–Hangzhou Grand Canal Basin in Changzhou City. The method begins by creating feature sequences for each site based on the time lag relationship of total phosphorus concentration between sites. It then constructs spatiotemporal graph data by combining the real river distance between sites and the correlation of feature sequences. Next, spatial features are extracted by fusing node features using the graph attention (GAT) module. The study employs the Informer network, which uses a sparse attention mechanism to extract temporal features efficiently for simulating and predicting total phosphorus data of the sites. The model was evaluated using R2, MAE, and RMSE, with the experimental results yielding values of 0.9619, 0.1489%, and 0.1999%, respectively. The GAT-Informer model exhibits enhanced robustness and superior predictive accuracy in comparison to traditional water quality prediction models.
7 January 2026 in Environmental Monitoring and Assessment
Organic pollution in the lake water bodies poses a serious threat to the stability of aquatic ecosystems and human health. Dissolved organic matter (DOM) is a key component of organic pollution. The analysis of its sources is crucial for pollution control. In order to accurately trace the source of organic pollution in the Changdang Lake basin, this study proposes a traceability method combining three-dimensional excitation-emission matrix (3D-EEM) fluorescence spectroscopy technology with a deep learning model. First, water quality samples were collected from the rivers connected to Changdang Lake and surrounding industrial wastewater, agriculture, and domestic pollution sources. Raw data were obtained by 3D-EEM spectroscopy. Parallel factor analysis (PARAFAC) was used to analyze the fluorescence data. Obtain fluorescence spectral images that characterize different pollution sources. Industrial wastewater, agricultural, and domestic pollution sources were used as training data (labeled 0, 1, and 2, respectively) to build and pre-train a deep learning model. Four deep learning models (Transformer, GoogLeNet, VGG, and AlexNet) were selected for comparison. Transformer model performs better in terms of both recognition efficiency and accuracy. The fluorescence spectrum images of the rivers connected to Changdang Lake and Lake body were input into the trained Transformer model to identify the sources of pollution. Tucker Congruence (TC) coefficients are introduced to quantify and verify the recognition results. The results show that for 40 fluorescent components in the Changdang Lake basin, the identification results of 38 components are consistent with the TC coefficients, with an identification accuracy rate of 95%. Compared with the traditional manual tracing method that relies on TC coefficients, this method significantly reduces the workload. The time required has been reduced from hours to minutes. This study provides efficient and reliable technical support for tracing the source of organic pollution in lake basins.