Time Series Prediction Of Water Quality In Water Source Area Based On LSTM Multi-point View Online Learning

As the rapid development of industrialization and urbanization,the water pollution situation of surface water mainly including lakes and rivers has become very serious.Especially the water pollution situation of urban water source is the most serious,urban water source is suffering from frequent pollution of all kinds of untreated sewage of the city.Water plant produce substandard water leading to water shortage in cities because of the deterioration of water pollution.Moreover,substandard water can directly endanger the life and health of urban residents.Therefore,we need to propose effective measures of water quality monitoring of water source area.There are many ways to monitor water quality,among which time series prediction is a common method.the overall trend of water quality changes in the future can be obtained by time series prediction.Relevant departments can manage the water quality of water sources based on the trend of water quality changes.However,it is not easy to obtain a suitable and accurate prediction model that can evaluate water quality across the board.Therefore,this paper conducts experiment on water resource area of a water plant in Tai hu Lake Jinshu Port of Suzhou.By setting five points in the monitoring area of the water source,each point collects six water quality indicators,including p H,DO,EC,turbidity,NH_3-N and COD.The collected data are preprocessed,analyzed and modeled before the experiment.The main contributions of this paper are as follows:（1）Traditional time series prediction models（such as AR,ARIMA,Grey prediction model,etc.）are limited by high-dimensional data,complex function representation and prediction scenarios.They cannot be used in these situation.So,machine learning and deep learning are selected to model water quality data.The research select XGBoost,Radial Basis Function neural network（RBF neural network）and Long short-term memory neural network（LSTM neural network）to realize end-to-end learning prediction of single-point water quality time series data.The results of the experiment show that LSTM is more effective.（2）Considering that the collection of data in the real data is not done immediately and water quality data is often in the form of data stream,conventional offline learning can not meet the needs of water quality time series prediction.On the basis of LSTM gradient learning way,the experiment realizes online learning and simulates data stream water quality data to predict water quality.According to the end-to-end prediction experiment of single-point water quality time series,the results of the experiment show that compared with the prediction effect of offline learning in（1）,LSTM online learning can greatly improve the prediction of water quality time series.（3）In view of the water quality data in the paper are multi-source data,water quality data at 5points are used to describe the overall water quality status of the water source region.Considering the overall analysis of data of water quality research area,this paper proposes a LSTM multi-point view online learning model.In this model,water quality data at each point is seen as a view data and an LSTM subnetwork is built for each view data respectively.The connection layer is used to connect each subnetwork and online learning is used for optimization.The experiment shows that the LSTM multi-point view online learning model can realize the monitoring task of the whole water quality in the source area.