Water Quality Data Analysis Based On Neural Network

Because the energy and forests are exploited and used as well as the waste gas and waste water discharge,the environment on which we live has been seriously impaired.The destruction of ecological environment has become one of the prominent contradictions restricting the sustainable development of our country.Although general control in key river areas has been enhanced,eutrophication in lakes characterized by nitrogen and phosphorus pollution is still a major problem to be solved in China.The research and prediction of water quality trends can afford a rational evidence for the comprehensive management and decision support of water environment quality.The main research contents of this dissertation are as follows:1.The periodicity and correlation of dissolved oxygen(DO),total nitrogen(TN)and total phosphorus(TP)from Taihu are studied by wavelet analysis,cross wavelet analysis and wavelet correlation analysis.The experimental results show that the DO has key periods,which are 200 month(M)and 110-125 M respectively.TN has a key cycle,90 M.Before 2003,it can be seen that the key cycle of the TP is 50 M respectively.However,because of the excessive intervention of human factors in 2007,the subsequent cycle is not obvious.The figures for TN and DO increase inversely.And the periodic correlation is strong at 1-5M and64 M.DO and TP show a regional variation in the period of 64 M,and DO is ahead of TP.The correlation between TN and TP is stronger in 8-16 M and 16-32 M,showing a positive correlation between the two,and TN is ahead of TP.2.The thesis bring forward a two-layer Gated Recurring Neural Network(GRU)model to forecast water quality.Firstly,the best prediction model is procured by comparing the forecast errors of different parameter models.In the prediction results of 6 months,5 years and 7 years,the errors of this model all rank at the top,whose maximum is 1.5,among which the Nash efficiency coefficient values are 0.645,0.73 and 0.692 respectively.Then,through comparing the Long short-term memory neural networks(LSTM)with same parameters,the research show that comparing the results of the forecasting models in different periods,the errors of GRU are smaller,and the difference of Mse is 0.9,whose minimum is 0.4.NSEC is higher,in 5 years and 7 years data to predict,whose difference of 0.25.In conclusion,the GRU model is better than the LSTM prediction model with the same parameters.3.GRU model requires manual parameter adjustment,which takes a long time.In order to solve this problem,this thesis bring forward an improved particle swarm optimization algorithm(PSO)for the parameter optimization of GRU.Experimental consequences show that comparing the prediction results of the GRU model with the improved PSO,the LSTM model and the auto-regressive differential moving model,the error values of the GRU model with the improved PSO are the smallest in each prediction period.The maximum error is1.593,and the minimum error is 0.914.Nash efficiency coefficient is biggest,whose maximum value is 0.848,and the minimum value of 0.719.In summary,the improved algorithm is the best.