Water Quality Prediction Method Based On BiLSTM And Transfer Learning

Water is the source of life.At present,the global water shortage is becoming more and more serious,and the deterioration of water quality is a prominent problem.Water quality prediction is one of the important means to carry out water pollution prevention and control.Efficient and accurate realization of the prediction of key water quality indicators and knowing the health status and trends of water bodies in advance is the basis for water quality assessment.It has practical significance for the effective use of production and domestic water and the early warning and forecast of water pollution,and plays a key role in water safety management and control,water resource allocation,and water supply systems.The specific work of water quality prediction is based on the monitoring data of water quality monitoring stations,using prediction models to predict the health status and changing trends of water quality.The authenticity and reliability of water quality data are the decisive factors for the predictive performance of the model.However,water quality monitoring stations exposed to the natural environment cannot avoid noise pollution,and large number of water quality indicators with significant differences in correlation and insufficient monitoring data also bring difficulties to water quality data prediction.In order to solve the problem of noise pollution in water quality data,after analyzing the water quality data,according to the characteristics of strong noise,non-stationarity and periodicity of the water quality data,the wavelet threshold denoising method is selected to denoise the data.Using signal-to-noise ratio,water quality data periodicity and water quality index hierarchical continuity test as evaluation indicators,the parameters of wavelet threshold denoising are determined through experiments.The denoising results have passed the water quality index hierarchical continuity test,ensuring the continuity of the water quality index time series in the water quality classification,and a good denoising effect has been achieved.The research object of this paper is two types of monitoring stations distributed along the same channel.Among them,type A stations have sufficient data,and type B stations lack sufficient monitoring data.For Type A monitoring stations with sufficient monitoring data,an NRS-Bi LSTM water quality prediction model was built.Using water quality index ratings with practical engineering value as decision-making attributes,and using neighborhood rough set theory to reduce the attributes of many factors that affect the target water quality indicators.And the dimensionality of the input data is reduced from the original 30 indicators to less than10 indicators,which simplifies the input of the model.Selected 5 indicators of 3 measuring stations in a certain water quality monitoring system for comparative experiments.The experimental results show that NRS-Bi LSTM can fully extract the time series characteristics of water quality data.Compared with the traditional neural network,the average error is reduced by 62.18%,and there is no lag in the prediction results,which shows better performance in water quality prediction.For Type B monitoring stations with insufficient monitoring data,an MMD transfer learning water quality prediction model was built.Considering the large number of water quality monitoring stations along the route,a migration strategy based on the distribution distance threshold is quantitatively given for the data distribution between the stations.The MMD algorithm is used to calculate the distribution distances of 12 Type A stations and 30 Type B stations,and the distribution distances at the top 5% and top 95% of the calculated results are used as the two threshold distances of the model.Using three distributed distance measuring stations as the source domain,a comparative experiment was carried out on the prediction effects of different migration strategies.The experimental results show that selecting the appropriate source area and migration strategy according to the distribution difference of the water quality data of the measuring stations can effectively improve the water quality prediction accuracy,and realize the water quality prediction of the B type measuring stations with insufficient data.