Prediction Of Water Quality In The South Source Of Qiantang River Basin Based On Single And Multi-step Models

Our country is faced with severe shortage of water resources and water pollution problems.Using scientific and reasonable methods to effectively supervise and control the water environment has become an urgent task to solve the water environment problems.Water quality prediction is a basic work of water environment treatment.Establishing an efficient and accurate water quality prediction model can not only directly reflect the pollution degree of water body,but also directly show the future trend of water quality items.Accordingly,it can timely adjust the relevant work of water pollution prevention and control,and provide a solid and reliable guarantee for water environment treatment.Based on the project of"Zhejiang Provincial Surface Water Quality Forecast and Early Warning System（ZJCT5-2020154）"of Zhejiang Provincial Ecological Environment Monitoring Center,taking the south source of Qiantang River basin in Zhejiang Province as the research area,and based on meteorological,hydrological and water quality data,this paper studies the single and multi-step prediction models of surface water quality,and realizes the single and multi-day forecast of p H,DO,COD_Mn,NH₃-N,TN and TP in all monitoring sections of the basin.The main research contents are as follows:（1）This paper firstly collects the input data needed for modeling,including basic information data and automatic monitoring data.Then,the outliers in the data set are detected by the 3σ criterion,the missing values are filled or replaced by the Lagrange interpolation polynomial method,and the data set after data integration is standardized by the maximum and minimum method.The correlation among meteorological,hydrological and water quality data,and the spatial correlation among water quality items of each monitoring section are analyzed.（2）Aiming at the problem of single-step water quality prediction,this paper firstly establishes two single-step water quality prediction models based on LSTM and BP,and then proposes a single-step water quality prediction model based on LSTM-BP by comprehensively utilizing the ability of LSTM to memorize long-term dependence and the nonlinear fitting ability of BP.Firstly,a time simulator based on LSTM is established for each monitoring section to extract the complex nonlinear relationship and time characteristics of meteorological,hydrological and water quality data of each monitoring section.Then,the spatial relationship between water quality items of each monitoring section is extracted by using the spatial combiner based on BP.Experimental results show that LSTM-BP model has the best prediction performance,followed by LSTM model,and then BP model,and these three models can provide a certain reference for the single-step prediction of the water quality items of each monitoring section in the basin.（3）Aiming at the problem of multi-step water quality prediction,based on the single-step water quality prediction models,this paper proposes a multi-step water quality prediction model based on BP using recursive strategy,a multi-step water quality prediction model based on attention encoder-decoder LSTM and a multi-step water quality prediction model based on attention encoder-decoder LSTM-BP using multi-output strategy.The multi-step water quality prediction model based on BP is realized by iterative single-step prediction model,while the attention encoder-decoder LSTM and the attention encoder-decoder LSTM-BP model are sequence-to-sequence models based on multi-output strategy.The multi-step water quality prediction model based on attention encoder-decoder LSTM extracts the complex nonlinear relationship among meteorological,hydrological and water quality data by constructing multi-variable input at the encoder end,extracts the dependent characteristics of time series data at different moments by using the attention mechanism,and extracts the influence of meteorological forecast data on multi-step water quality prediction value by connecting the context vector obtained by attention mechanism with meteorological data at k time in the future.The multi-step water quality prediction model based on attention encoder-decoder LSTM-BP uses a spatial combiner based on BP to extract the spatial relationship between water quality items in each monitoring section,on the basis of establishing a time simulator based on attention encoder-decoder LSTM for each monitoring section.Experimental results show that the prediction performance of each model is gradually deteriorating with the increase of prediction steps,but the prediction performance of the attention encoder-decoder LSTM-BP model is best at all steps,followed by attention encoder-decoder LSTM model,and then BP model,and these three models can provide a certain reference for the multi-step prediction of the water quality items of each monitoring section in the basin.（4）This paper constructs a surface water quality prediction system,which can realize the functions of user login,data acquisition,data storage,data processing,model design and data display.The system can automatically obtain the daily data of each meteorological station,hydrological station and monitoring section in the basin,and automatically save them in the My SQL database table.Through the web page,it can directly observe the historical data change trend,the single and multi-step prediction results of each water quality prediction model of each water quality item of each monitoring section in the basin.