Research On Prediction Of Total Phosphorus In Rivers Based On Multi-graph Convolutional Network

To monitor the current state of the water environment as well as future trends,water quality prediction has a unique role and significance.Among water pollution,river pollution is of great concern because of its rapid spread and high pollution impact.Since there are many factors affecting river pollution and the pollution sources are distributed in a faceted manner,it makes it difficult to predict water pollution.For the problems that the existing river water quality prediction methods in the spatial correlation factors are not enough to consider,and the river water quality between upstream and downstream monitoring stations has a time asynchronous response effect,this thesis researches the prediction method of total phosphorus in rivers based on multi-graph convolutional network,which selects Pengshan District in the Middle Reaches of Minjiang River as the study area and uses one of the main indicators of river water quality,total phosphorus as the study object.The details are as follows:（1）In order to analyze the influencing factors of total phosphorus in rivers,the analysis of spatial and temporal variation characteristics and driving mechanism of total phosphorus in rivers in Pengshan District was carried out.The analysis results show that there is good consistency in the temporal variation of total phosphorus in rivers in Pengshan District,with overall high total phosphorus concentration in spring and summer as well as low concentration in autumn and winter,but showing both some spatial correlation and spatial variability in space,among which the correlation is stronger in neighboring areas.The correlation analysis between meteorological factors,human activities and river total phosphorus in Pengshan District show that total phosphorus is moderately correlated with temperature and barometric pressure,and weakly correlated with total precipitation,evaporation and relative humidity,while there is a significant correlation between land use patterns and total phosphorus.（2）To comprehensively consider the temporal autocorrelation and spatial interdependence of total phosphorus in rivers,this thesis combines multi-graph convolutional network and temporal convolutional network to propose a river total phosphorus prediction model based on MGCN-TPA-TCN.Aiming at the problem of temporal asynchronous response of total phosphorus between upstream and downstream river monitoring stations,a temporal convolutional network river total phosphorus prediction model based on temporal pattern attention mechanism（TPA-TCN）is proposed.And the TPA mechanism can differentially assign attention weights at different time steps,which in turn enhance the prediction performance of the model.To address the problem that river total phosphorus prediction has insufficient consideration of spatial correlation,this thesis constructs multiple relationship graphs to fully exploit and utilize the spatial dependence of river total phosphorus.And then multiple spatial features of total phosphorus are captured and fused by a multi-graph convolutional network,and the temporal features of total phosphorus are captured by a TPA-TCN model,which can further improve the model prediction accuracy.（3）In order to test the applicability of river total phosphorus prediction model based on MGCN-TPA-TCN in rivers of Pengshan District,through training experiments with a large number of measured data,the best hyperparameter configuration was screened in this thesis to achieve an effective prediction of river total phosphorus in Pengshan District.Through a variety of comparative experimental analyses,the results show that the prediction accuracy of the MGCN-TPA-TCN model,which incorporates multiple spatial relationship graphs,is superior to other prediction models,with the lowest RMSE value reaching 0.0102 mg·L^-1,MAE value of 0.0068 mg·L^-1,and MAPE value of 3.42%,which reduces RMSE by 44.57%,MAE by 41.38%and MAPE by27.25%compared with the TCN model.This study can provide methodological reference for river water quality prediction.