基于GCN-LSTM的钱塘江南源水质预测研究.

The water quality data of the river basin has dual dependencies in time and space. Based on the fact that most of the existing water quality prediction models are only based on the time dimension and cannot make effective use of the spatial correlation of water quality, a spatiotemporal prediction model based on graph convolution neural network GCN and long short-term memory network LSTM was proposed in this paper. Firstly, the whole watershed was constructed into the topology of station-to-station connection, and the distance along the river between stations was taken as the indicator of connectivity between stations. For each input time, an independent GCN was used to capture the spatial relationship between stations, and then the temporal variation characteristics were captured by LSTM. Finally, the multi-step prediction results were obtained by multi-layer perceptron (MLP). The south source of Qiantang River was selected as the research object, multi step prediction experiments were carried out on three water quality prediction tasks of pH, DO and CODMn at 15 monitoring stations in the basin. The results show that in the three water quality prediction tasks, the average percentage error MAPE of GCN-LSTM model compared with LSTM model is decreased by 15.29%, 11.77% and 9.8% respectively. Compared with some classical time series models, the model proposed in this paper can better capture the spatial relationship of watershed water quality, improve the accuracy of water quality prediction and provide a new method for watershed water environment quality prediction. The connections between monitoring points have differences, and using distance to represent the strength of the connections between monitoring points can make water quality prediction results more accurate. [ABSTRACT FROM AUTHOR]