An improved graph neural network integrating indicator attention and spatio-temporal correlation for dissolved oxygen prediction

Abstract

Accurately predicting dissolved oxygen (DO) is essential for water environment protection and management. The spatiotemporal dependencies of water quality and the interactions between indicators are neglected in existing prediction models. To improve the DO prediction accuracy, a graph neural network based on indicator attention mechanism and bayesian optimization (BO-AM-MTGNN) was proposed in this study. Hourly water quality data at 20 sampling sites in the Chaohu Lake basin from January 2022 to February 2024 were used as the research dataset. The effectiveness of the BO-AM-MTGNN model was validated through comparisons with baseline models (XGBoost, LightGBM, LSTM, GRU, Informer) and ablation experiment (BO-AM-MTGNN, AM-MTGNN, MTGNN). The results demonstrated that the BO-AM-MTGNN model effectively captured the temporal and spatial information of water quality data. Correlations between indicators can be fully extracted by the indicator attention mechanism. Compared with the MTGNN model, the MAE, RMSE, and MAPE of the BO-AM-MTGNN model decreased by 12.16 %, 5.50 %, and 12.13 %, respectively. The prediction accuracy of MTGNN outperformed the baseline models, with the performance ranking as follows: MTGNN > Informer > LSTM > GRU > LightGBM > XGBoost. The BO-AM-MTGNN model proposed in this study effectively improves DO prediction accuracy. In future studies, the BO-AM-MTGNN model holds potential for water quality early warning and pollution source tracking.