Sponge City Runoff Prediction Data Driven Model Coupling Variable Processing Algorithm

With the acceleration of urbanization,the problem of urban rainwater management has become increasingly prominent.In order to alleviate a series of problems caused by urbanization,such as urban waterlogging,shortage of water resources,and water pollution,China proposes to build a"sponge city"with natural accumulation,natural infiltration,and natural purification to restore natural water circulation.During the construction of sponge cities,rigid indicators such as annual runoff control rate and runoff pollution reduction rate are usually used to evaluate and predict their construction effectiveness.However,the current estimation of runoff and runoff pollution in sponge cities relies on mechanism-based distributed hydrological models of the watershed or a large number of online monitoring facilities,and there is a lack of simple and efficient evaluation and prediction tools.At present,the commonly used urban watershed hydrological models at home and abroad are mainly based on the US EPA SWMM and a series of commercial software derived from them.They are mechanism models and have the fundamental characteristics of"process-driven".There are many complex parameters,causing model calibration difficult.They also need professional expertise for technical personnel,which restrict their promotion.With the continuous development of high technology such as big data and cloud computing,artificial intelligence models supported by online monitoring big data have become the development trend of urban rainwater models.Among them,the data-driven model represented by artificial neural network(ANN)has been applied in simulating natural river hydrology due to its good ability to construct complex nonlinear mapping.With the vigorous promotion of the sponge city construction model in our country,the need for rapid and accurate prediction of sponge city construction indicators is becoming increasingly urgent.In response to the major technical requirements that frequently need to assess and predict the completion of total runoff control,runoff peak control,runoff pollution control and other indicators during the construction of sponge cities,the national-level sponge city pilot Chongqing Yuelai New Town International Expo Center Area is studied,based on the measured runoff data of 8 monitoring stations from 2019 to 2020.A data-driven model has been carried out to evaluate and predict the runoff of watersheds with runoff control facilities.This research explores the variable preprocessing methods of stepwise regression analysis and principal component analysis,explores the coupling mode between variable preprocessing algorithms and data-driven models,and constructs and studies the impact of multiple linear regression and BP neural network on the areas with sponge facilities.Forecast performance and accuracy of total runoff,peak flow and pollution load.The study uses factors such as rainfall factors,underlying surface conditions,and scale of low-impact development facilities as the original variables,and uses stepwise regression analysis to screen the original variables to construct a multiple linear regression of runoff prediction,runoff peak flow prediction,and runoff pollution load prediction.Principal component analysis is used to reduce the original variables to construct a BP neural network model for total runoff prediction,runoff peak flow prediction,and runoff pollution load prediction.A variety of comprehensive indicators are used to evaluate the effectiveness of the two types of models and compare the two types of models.The main research contents and conclusions are as follows:(1)A multiple linear regression model and a BP neural network model for the prediction of the total runoff of sponge citiesThrough stepwise regression analysis,the original variables that have a significant impact on the total runoff were selected to construct a multiple linear regression model for predicting the total runoff.Principal component analysis was performed on the original variables,and the six principal components obtained from the principal component analysis were used as input variables to construct a BP neural network model for total runoff prediction.According to the prediction results of the test set,the multivariate linear regression model R~2 of the total runoff is 0.6208;the BP neural network model R~2 of the total runoff prediction is 0.8475.Compared with the multiple linear regression model,the BP neural network model increased the value of NSE by36.6%,reduced the value of RMSE by 13.5%,and increased the value of R~2 by 36.5%.The results show that the BP neural network can more effectively predict the total runoff of sponge cities.(2)A multiple linear regression model of the peak flow of sponge city runoff and a BP neural network modelSimilar to the sponge city total runoff prediction model,a multivariate linear regression model for the prediction of peak runoff flow was constructed by using stepwise regression analysis to select variables.The six principal components obtained by principal component analysis construct a BP neural network model for the prediction of peak runoff flow.In the test phase,the multiple linear regression model R~2 of the peak runoff flow is0.4505;the BP neural network model R~2 of the peak runoff flow is 0.8078.Compared with the multiple linear regression model,the BP neural network model increased the value of NSE by 71.3%,reduced the value of RMSE by 38.8%,and increased the value of R~2 by 79.3%.The BP neural network model of peak runoff flow has been significantly improved in terms of prediction effect.(3)A multiple linear regression model and a BP neural network model of the sponge city runoff pollution loadFirst,the event mean concentration(EMC)of pollutants in rainfall runoff was calculated based on the original monitoring data.According to the screening results of stepwise regression analysis,a multiple linear regression model of runoff pollution load was constructed.The BP neural network model of runoff pollution load is constructed by the six principal components obtained by principal component analysis.The R~2 of the runoff pollution load multivariate linear regression model test set was 0.5557,and the R~2of the runoff pollution load BP neural network model test set was 0.8161.Compared with the multiple linear regression model,the BP neural network model increased the value of NSE by 41.7%,reduced the value of RMSE by 35.8%,and increased the value of R~2 by46.9%.The results show that the BP neural network also has a good effect in the prediction of runoff pollutants.In general,the BP neural network model shows better prediction performance for the three key indicators of sponge city construction effectiveness evaluation of total runoff,peak runoff flow,and runoff pollution load.Since the rainfall-runoff relationship is one of the most complex hydrological phenomena in nature,and multiple linear regression has a relatively good effect in capturing the linear relationship,the BP neural network have better results for the three types of models with better fitting ability for the nonlinear relationship.In the process of planning and designing specific runoff control facilities,using relevant models to evaluate and predict their runoff control effects is a prerequisite for ensuring the scientific and economical construction of sponge cities.The two data-driven model frameworks constructed in this study provide low-cost and high-precision calculation tools for the construction of sponge cities and the design of runoff control facilities in newly developed areas with similar geographic conditions,allowing decision-makers and planners to make more scientific and accurate planning and layout of green infrastructure for runoff control at the beginning of the overall planning of the watershed.