IEEE Access (Jan 2024)
Design of Prediction Framework Geo-TA Utilizing Spatial and Temporal Water Quality Data Integrating Meteorological Information
Abstract
Water quality predictions are crucial for water resource management and aquatic biological protection.Current prediction methods typically rely solely on temporal water quality data from a limited number of monitoring stations. This could lead to inaccurate predictions due to the lack of other data sources. To address this issue, we propose the Geo-TA (Geographical-Temporal Attention with Cross-Attention Mechanism) framework, based on GATnet, which incorporates the geographical coordinates of water quality stations and uses meteorological data as additional predictors. In the Geo-TA framework, datasets from the Southeast, Western, and Northern regions, collected from 18 water quality monitoring stations along the Liaohe River in Liaoning Province, China, are utilized to predict multiple water quality indicators, including WT, pH, DO, CODMn, NH3-N, P and N. We used the Val-Huber Loss of the Geo-TA framework to compare with that of the baseline models such as LSTM-CNN-ATT, MLP and TCN in predicting accuracy. In the scenario of low concentrations of pollution, the validation Huber Loss values of the Geo-TA framework are 0.014198 on the Southeast dataset and 0.055576 on the Western dataset. LSTM-CNN-ATT performs the worst on the Southeast dataset with a validation Huber Loss as high as 0.821015, while TCN performs the worst on the Western dataset with a validation Huber Loss as high as 0.699873. The validation Huber Loss of Geo-TA is consistently lower than that of the three baseline models. In the heavily polluted Northern dataset, all three baselines exhibit poor performance, while Geo-TA shows excellent predictive capabilities, with the Val-Huber Loss of only 0.021582. The Geo-TA framework achieves higher prediction accuracy than the baseline models by comprehensively considering water quality data, spatial information, and meteorological data.
Keywords
