Design of a Variable-Length Sequential Prediction Framework GTV-STP Based on Spatial and Temporal Water Quality Information of Taihu Lake

Water quality predictions have a great importance in water resource managements and water pollution protections. Most the currently used water quality models can only predict time sequence of fixed length and ignore the real hydrological flow information. In this paper, we propose a new water quality prediction framework GTV-STP that adopts an embedding-Encoder-Decoder structure, in which the timestamps are introduced into the water quality information. The spatial embedding method is developed to introduce the spatial hydrological features into the water quality information, in order to achieve multisite parallel predictions. In addition, a variablelength decoder is put forward for the variable-length sequencial predictions. The practical water quality predictions of WT, pH, DO, CODMn, NH3-N, P and N for Taihu Lake with GTV-STP framework are performed on two datasets of 6to12 in smaller-scale and 12to24 in larger-scale. Using the Huber Loss to balance the MAE and the RMSE, the val-Huber Loss of the GTV-STP framework is used to compare with that of the baseline models such as LSTM-CNN-ATT, LSTMNet and MLP in predicting accuracy. Results show that the GTV-STP framework has the highest accuracy with the val-Huber Loss of 0.079183 on 6to12 and 0.033561 on 12to24. It is shown that the GTV-STP framework has highly accurate not only for smaller-scale water quality predictions but also suitable for larger-scale predictions. In future, water quality predictions using other more precise frameworks containing spatial and time series information under deep learning will be one of the research directions.