BP Neural Network Based Pollution Control Assessment Model For Stormwater Constructed Wetlands And Its Integration With SWMM

Urban runoff pollution has many negative environmental impacts on urban receiving water bodies.With the gradual transformation of rainwater management measures from engineered to ecological,constructed wetlands were increasingly used to intercept urban runoff pollutants.At the beginning of the design of stormwater constructed wetlands,it is the prerequisite to ensure the scientific and economic design of stormwater constructed wetlands by using the relevant model to predict the runoff treatment performance according to the pollution load of its service area.At present,there are some shortcomings in the prediction model of the pollutant purification ability of constructed wetlands such as complex parameters,low accuracy and lack of linking to pollution loads from dynamic rainfall processes.A BP neural network model was established in this study.This study used different variable selection methods(principal component analysis(PCA),stepwise regression analysis(SRA)and mean impact value analysis(MIVA))to select the most important and representative variables from many factors influencing the pollutant purification performance of constructed wetlands as input variables,and established the BP neural network prediction model of effluent SS,TCOD,TN and TP concentration.On this basis,a BP neural network model with steep slope features of mountain cities incorporated was constructed,and the effects of cascade inlets on SS,TCOD,TN and TP removal performance of stormwater constructed wetlands in mountain cities were examined.Considering the lack and uneven distribution of data,Markov correction was carried out to modify the prediction results of BP neural network model to improve the overall prediction performance using Markov chain model.The BP neural network model was coupled with the SWMM model.SWMM model was used to provide the input value of urban nonpoint source pollution load under dynamic rainfall for BP neural network model.The hybrid model was constructed and applied in a practical engineering project.The research results provide a tool for assessment of the runoff pollutant reduction performance at constructed wetlands and watersheds level,and provide ideas for the design and performance prediction of other runoff pollutant control facilities.The main research content and conclusions were as follows:(1)Selection of influence factors and development of pollutant removal performance prediction model for stormwater constructed wetland.PCA,SRA and MIVA were used to select the most representative influence factor of runoff pollutant removal in stormwater constructed wetlands,which can be used as input variables to construct the pollutant reduction prediction model based on the artificial neural network model and BP algorithm.The prediction accuracy of the model constructed by different variable selection methods were compared to examine the influence of different variable selection methods on the model prediction performance.The prediction performance of BP neural network model and traditional multi-linear regression model were compared to verify the superiority of BP neural network model.1)In the process of constructing the prediction model of runoff pollutant removal performance in stormwater constructed wetlands,the prediction performance of models using the most representative variables obtained by variable selection methods is generally better than that of models constructed by using all the original variables.However,the variable selection method and the most representative variables groups show obvious different response characteristics for different pollutant types.For the effluent SS,TCOD,TN and TP concentration prediction of constructed wetlands treating stormwater runoff,the best variable selection methods were MIVA,SRA,MIVA and PCA,respectively,when adopting single variable selection methods,the R~2of the corresponding model were 0.983,0.823,0.978 and 0.985.For the effluent SS,TCOD,TN and TP concentration prediction of constructed wetlands treating stormwater runoff when using multiple variable selection methods,the best variable selection methods were SRA∩PCA,SRA∩PCA∩MIVA,SRA∩MIVA and SRA∩MIVA,respectively,the R~2 of the corresponding model were 0.919,0.625,0.992 and 0.962.2)Based on performance development between the BP neural network model and the multiregression model,it was found that the performance of BP neural network model is better than that of the multiregression model in general.For the effluent SS,TP concentration prediction of constructed wetlands treating stormwater runoff,the difference between the prediction performances of these two kinds of models was not large.The prediction performance of MLP models could be better if a larger dataset could be used.However,it can be seen from the prediction of the effluent TCOD,TN concentration that the MLP neural network model has an advantage over the MLR model.(2)BP neural network model construction of water quality prediction of constructed wetlands treating stormwater runoff considering step characteristic of mountain cities.Considering that cascade structures were widely used in biodetention facilities and stormwater constructed wetlands in mountain cities,the cascade inlet width and height were selected as factors to characterize the characteristics of mountain ladders.Two BP neural network models containing and not including mountain ladder features were constructed,and the prediction performances were compared to study the effects of inlet width and height on the pollutant removal of stormwater constructed wetlands.1)For the effluent SS,TP concentration prediction of constructed wetlands treating stormwater runoff,the MLP models considering step characteristic of mountain cities perform better.The R~2 of constructed models were 0.975 and 0.982,respectively.This shows that the characteristic of cascade inlets of constructed wetlands treating runoff has an impact on SS,TP reduction.2)For the effluent TCOD,TN concentration prediction of constructed wetlands treating stormwater runoff,the MLP models not considering step characteristic of mountain cities perform better.The impact of characteristic of cascade inlets of constructed wetlands treating runoff on TCOD,TN reduction is not found.The runoff in mountain cities has a faster flow rate due to the terrain,so dissolved oxygen concentration in the stormwater runoff is relatively high.Hence,the impact of the reaeration effect of cascade inlets of constructed wetlands on pollutant reduction is weakened.Therefore,in the prediction model development of effluent SS,TP concentration of constructed wetlands treating runoff,it is not recommended to include the width and height of the"falling water"inlet into the input variables.(3)Markov correction of BP neural networks for water quality prediction of constructed wetlands treating stormwater runoff.In view of the potential lysinity of monitoring data and the imbalance of data distribution,the coupling of the statistical Markov chain(Markov)model is corrected to correct the BP neural network model.For the effluent SS,TCOD,TN and TP concentration prediction of constructed wetlands treating stormwater runoff,the results of Markov correction were closer to the real values than the MLP prediction results.The Markov chain model can effectively increase the prediction accuracy of BP neural network.This shows that the BPNN-MC model combining the BP neural network and the Markov chain model has better reliability and prediction accuracy.(4)The development of SWMM-BPNN model for runoff pollutant reduction assessment in mountain city basins.The BP neural network model was coupled with the SWMM model.SWMM model was used to provide the input value of urban nonpoint source pollution load under dynamic rainfall for BP neural network model.The hybrid model was constructed and applied in Yuelai New Town in Chongqing,the water quality prediction model of BP neural network of constructed wetlands was coupled with SWMM,and the assessment of the reduction of runoff pollutants by the riverside terraced wetland group in this area was carried out.Using the SWMM’s runoff generation and concentration calculation results as pollutan loading inputs,the BP neural network evaluation model can be used as an efficiency prediction tool for specific runoff pollutant control facilities,realize the dynamic and accurate evaluation of the efficiency of runoff pollutant control facilities,and improve the accuracy of SWMM’s prediction efficiency for runoff control facilities.