Research On Prediction Method Of Spatiotemporal Water Quality Based On Graph Convolution

Water resources management is of great significance to urban daily water consumption,agricultural irrigation,industrial water consumption,and ecological environmental protection.The pollution of surface water resources in my country is still serious,and it is necessary to implement water pollution prevention and control actions.Therefore,it is necessary to grasp the possible future trends of water quality,and then establish a perfect water environment water quality prediction and early warning system.In response to this problem,this work proposes an integrated water quality spatiotemporal prediction model,which can combine historical spatio-temporal multi-factor characteristics to predict changes in water quality over a period of time in the future,and provide auxiliary decision-making for water environment management decisions.The work in this article mainly reflects the following two aspects of work,as follows:The first is the generation and processing of water quality time series.Due to the sensitivity of data in the field of water environment and backward data collection methods,it is difficult to obtain enough data in a short time to train the model to achieve water quality prediction.Therefore,it is necessary to use the existing water environment data to reasonably generate real and effective data,increase the amount of existing data,and thus improve the learning efficiency of the prediction model.But this is also a very challenging problem,affected by many complex factors.First,how to reasonably simulate the high-dimensional distribution of real data,rather than simply repeating the real data,is the primary problem in generating data.Secondly,for the generated simulation data,how to judge the pros and cons of the data simulation from different angles lacks a quantifiable evaluation index.Therefore,the use of objective and fair standards to measure the difference between the simulated data generated by different time series is also another problem faced by the current generated data.For these challenges,this paper comprehensively adopts the Generative Adversarial Network,learns through the game of the contest between the generative network and the discriminative network,and combines the multi-dimensional characteristics of the water environment to generate realistic simulation data to support the prediction research on the overall water environment status of the water environment management platform.Then compare several different prediction algorithms.This work selects ARIMA,Recurrent neural network,Graph convolutional network and single hidden layer neural network algorithms to model and predict water quality spatio-temporal multi-factor data.Based on the prediction accuracy performance of different models,an optimal water quality time series prediction model is obtained.Current water quality time series research generally only predicts the water quality transformation trend based on the time series angle of a single water quality station.This paper is based on graph convolution neural network.In addition to considering the water quality time series characteristics,it also further considers the spatial characteristics of water quality and comprehensively considers time and space.Multi-element features predict water quality,which can effectively improve the prediction accuracy of the model.In summary,this work is to complete the missing water quality data,and after the feature screening of the multi-element data,use the time-space-based multi-element graph convolution model to predict the water quality data.After experimental verification,the water quality time series prediction model based on the convolution of the space-time map has a greater improvement in accuracy compared with the traditional time series prediction model,and provides reliable data guidance for water environment management decision-making.