Research On Optimization Of Disinfection Dosing In Water Plants Using Bagging Heterogeneous Integrated Combination Prediction Model Based On RBF And SVM

At present,the dosage of disinfectants in most water plants in China mainly depends on manual experience,which has hysteresis and uncertainty.Especially in the case of sudden changes in raw water quality and flow rate,too little dosage of disinfectants can affect the water quality of the effluent.Excessive dosage not only increases production costs,but also increases disinfection by-products,affecting water quality safety.With the improvement of people’s quality of life,the development of information digitization,and the continuous expansion of application fields,the prediction of disinfection dosage after filtration in water plants is facing both opportunities and challenges.As an important link in the orderly development of the social economy and the protection of people’s lives and health,the disinfection of tap water has entered a period of transformation from"extensive"to"high-quality"development,and from"traditional mode"to"intelligent mode"in urban water affairs.Faced with increasingly high water quality requirements and exponentially increasing water supply demand,operators need to spend more time and resources monitoring water quality to ensure water quality safety.How to more accurately control the dosage of disinfectants,ensure urban water quality safety,and improve the management ability of water enterprises has become a common concern for water plants.Therefore,researching an effective integrated combination prediction model for disinfection dosage after filtration in water plants to improve the stability and accuracy of disinfection dosage prediction is an important topic for intelligent water management.Avoiding blind dosing of chemicals,reducing water production costs,and ensuring the safety of factory water quality are of long-term and important significance.In this context,based on the measured data of a water plant in Kunming,this study studied an integrated optimization algorithm based on support vector machine（SVM）and radial basis function neural network（RBF）models,and conducted comparative analysis and simulation predictions.The main work of this paper is summarized as follows:（1）Correlation analysis was conducted on 11 influencing factors,including turbidity（NTU）,residual chlorine(Cl_into),flow rate(Q_into),dosage of disinfectant,residual chlorine(Cl_out),oxygen consumption(COD_Mn),p H value,algae amount in water,ammonia nitrogen content（NH₃-N）,iron,manganese ions,and temperature.NTU,Cl_into,Q_into,Clout,NH3-N were selected The six influencing factors of COD_Mn are used as input factors for the integrated optimization prediction model of disinfection dosing,laying a foundation for the subsequent model establishment.（2）Given the Expansion factor（σ）And kernel function parameters in SVM（σ）、The penalty factor（C）is randomly generated and tends to fall into a local optimal solution during the prediction process.And has a good global optimization ability.Therefore,this paper uses particle swarm optimization（PSO）,random walk algorithm（RW）,and whale algorithm（WOA）to optimize RBF and SVM,and obtains six optimized prediction models（PSO-RBF,PSO-SVM,RW-RBF,RW-SVM,WOA-RBF,and WOA-SVM）.Through simulation testing of the optimized prediction model with measured data,it was found that the relative average errors of PSO-RBF,RW-RBF,WOA-RBF,PSO-SVM,RW-SVM,and WOA-SVM were 6.84%,3.79%,6.85%,5.32%,7.28%,and 7.88%,respectively,and the determination coefficients were0.51,0.84,0.51,0.70,0.53,and 0.39,respectively.The root mean square errors were 2.14,1.24,2.17,1.68,2.11,and 2.41,respectively.Compared to RBF,PSO-RBF,RW-RBF,and WOA-RBF,The average relative error decreased by 2.53%,5.58%,and 2.52%,the determination coefficient increased by 0.39,0.73,and 0.39,and the root mean square error decreased by 0.77,1.66,and 0.75,respectively.Compared to SVM,the average relative error of PSO-SVM,RW-SVM,and WOA-SVM decreased by 2.99%,1.03%,and 0.43%,respectively,and the coefficient of determination increased by 0.64,0.47,and 0.33,respectively.The root mean square error decreased by 1.31,0.88,and 0.58,respectively.In summary,RW-RBF and PSO-SVM have better prediction performance compared to other optimized prediction models.（3）Due to the large time-varying impact factors of the water plant,the anti-interference ability of the optimized prediction model is weak and the prediction accuracy is poor.In order to further improve the stability and prediction accuracy of the model,an integrated optimization prediction model（Bagging-PSO-SVM,Bagging-RW-RBF,and random forest）was established.Through simulation testing,it was found that the relative average errors of Bagging-RW-RBF,Bagging-PSO-SVM,and random forest were 3.60%,6.52%,and 8.58%,respectively,with determination coefficients of 0.89,0.86,and 0.53,root mean square errors of 1.46,2.34,and4.40,and residual prediction residuals of 2.59,2.08,and 1.46,respectively.Compared to PSO-SVM,the average relative error of Bagging-PSO-SVM decreases by 1.2,the determination coefficient increases by 0.16,and the root mean square error decreases by 0.66;Compared to RW-RBF,the average relative error of Bagging-RW-RBF decreased by 0.19%,the coefficient of determination increased by 0.05,and the root mean square error decreased by 0.22.In summary,the integrated optimization prediction model has significantly improved the prediction accuracy compared to the optimization prediction model,and Bagging-RW-RBF has better stability and prediction accuracy among the three integrated optimization prediction models.（4）In order to verify that Bagging-RW-RBF can still maintain good prediction accuracy under different data samples,500 and 800 sets of measured sample data were used for simulation verification of Bagging-RW-RBF.The relative average errors were 3.64%and3.31%,the determination coefficients were 0.85 and 0.78,the root mean square errors were1.05 and 1.12,and the remaining prediction residuals were 2.71 and 2.37,respectively.In summary,Bagging-RW-RBF can still maintain good prediction accuracy and stability under different data samples.