Research On Weak Time-delay Derived Neural Network Model For Determination Of Coagulant Dosing In Drinking Water Facility

With the continuous advancement of the automatic transformation of drinking water plants,on the premise of meeting the requirements of effluent quality and quantity,it has become a hot topic to explore more intelligent control methods to replace the traditional manual operation,reduce production energy consumption,reduce the operation cost of water plants and improve production efficiency.The coagulation process is the key step of water purification.The accurate control of coagulation dosing is the core content of automatic production and saving reagent cost in drinking water plants.Based on GA-BP neural network model,the dosage is predicted through raw water turbidity,temperature,p H,TDS,TP,UV254,flow and precipitation water turbidity.Using the method of time-delay measurement of precipitated water,the weak time-delay data that approximately replaces the effluent quality is obtained,the fitting performance of the weak time-delay GA-BP neural network model is analyzed,the factor sensitivity and influence factors of the weak time-delay GA-BP neural network model are explored,and the basis for model simplification is obtained.The main conclusions are as follows:(1)In the quarterly variation of raw water quality,turbidity,TP and UV254 have similar variation trends.The values increase significantly from May to June,and are relatively stable in other months;The temperature is mainly affected by season,and the measured temperature range is 10 ～ 35 ℃;The value of p H is relatively stable and always within the optimal p H range of flocculant;TDS decreased from May to June and remained relatively stable in other months.Among the daily changes of raw water quality,the changes of turbidity,TP and UV254 are more obvious,which are mainly disturbed by the change of intake flow.It increases with the increase of flow and decreases with the decrease of flow.The temperature,p H and TDS are relatively stable in a single day.In the relationship between factors and dosage,temperature,p H,turbidity,flow,TDS,TP and UV254 have different degrees of positive and negative correlation with dosage,but it is not a simple linear relationship.In the principal component analysis,the weight of inquiry factors in the comprehensive index,the effects of temperature and p H are similar,the effects of raw water turbidity,TP and UV254 are similar,and the effects of TDS,flow and precipitation water turbidity are different.(2)Among the time-delay and weak time-delay network models based on single factor and comprehensive factor,the prediction effect of single factor weak time-delay network model is the best.The correlation coefficients R2 of simulated and measured values fitted by training group and test group are 0.72 and 0.71 respectively,P values are 0.955 and 0.972 respectively,RMSE are 17.74 and 20.14 respectively,RMSE% are 16.79 and 19.17 respectively,pbias are 0,EF are 0.72 and 0.71 respectively,and D are 0.92 and 0.91 respectively.In the sensitivity analysis of single factor weak time-delay network model,the sensitivity of raw water turbidity is the largest,which is 0.794,followed by flow and temperature,which are 0.17 and 0.124 respectively.The sensitivity between TDS and raw water turbidity is negative,indicating a certain negative correlation with dosage.In the factor impact analysis,temperature,raw water turbidity,TDS,turbidity of precipitated water after one hour and flow have different effects on the weak time-delay network model,and the effects of p H,TP and UV254 on the model can be ignored.In the simplification of the model,the weak time-delay network model established by using five factors: temperature,raw water turbidity,TDS,sedimentation water turbidity after one hour and flow has no obvious difference between the prediction effect and the weak time-delay network without deleting factors.