Optimization Of Rainfall Runoff Data Driven Model Based On Deep Learning And Data Mining

As a non-engineering method,urban rainfall runoff simulation can not only assist planning,design and reform in the construction stage of the project,but also provide emergency simulation,decision support,operation and maintenance management in the operation stage of the project.To predict urban waterlogging risk quickly and accurately,it is important to construct model with high precision,high generalization and strong adaptability to local hydrological conditions.At present,the simulation of urban rainfall runoff mainly uses process-driven rainfall runoff models.The process-driven models have complex modeling conditions and their parameters are difficult to adjust.And the existing research methods are also difficult to improve the generalisability of the model parameters.In recent years,with the development of smart water construction in China,more and more cities are using sensors and Io T technology to monitor the parameters of urban water systems,which accumulates a large amount of effective data for urban rainfall and flood management.Therefore,it is worth exploring how to make full use of this data and mine it for valuable information.It is a worth exploring topic that builds a highly accurate,robust and generalisable data-driven model for predicting urban rainfall runoff based on data mining techniques and deep learning theory.This paper selects the Yue Lai Sheng Tai city in Chongqing Liangjiang New Area as the study area.Based on monitoring data from the"Yuelai Sponge City Monitoring and Information Platform",the adaptability of data-driven model is discussed.And the performance of the data-driven model is verified by comparing the prediction accuracy,operation efficiency and modeling efficiency of the process driven model SWMM and the constructed data driven model.The main results of this study are as follows:(1)Aiming at the problems of small data scale,lack of automatic calibration module and difficult data interaction in SWMM used rainfall runoff simulation.SWMM automatic parameter calibration is realized by coupling SWMM and genetic algorithm based on Python and Py SWMM.In order to ensure the generalization performance of the model,37 rainfall events were used to calibrate the model parameters during the training period,and 6 rainfall events were used to verify the model during the test period.Although the generalization performance of the model is ensured when SWMM uses large-scale data,the prediction effect of the model is not good.And only the NSE of 4 rainfall events is greater than 0.5 in the test set.(2)In order to explore the adaptability of data driven model in urban rainfall runoff simulation.Based on deep learning and optimization algorithm theory,this study constructs 4 data-driven models.Aiming at the influence of model structure on prediction results,grid search method is used to find the best structure of 4 models for this dataset.Then,all kinds of optimal structural models for training and testing,and compared with SWMM.The experimental results showed that the overall prediction performance of the4 data-driven models is stronger than that of SWMM for the 6 rainfall events in the test set.The improvement rates of RMSE,NSE and PE are obvious.LSTM is the best in the data-driven models.In the same iteration number,the running efficiency and modeling efficiency of the data-driven models are 6?23425 times and 182?7822 times of SWMM,respectively.But,same as SWMM,the four data-driven models also show better prediction performance for rainfall events with forward peak position,single peak type and short and moderate rainfall duration.(3)In view of the problem that the constructed data-driven model is still sensitive to the feature distribution of data samples,this paper further characterizes the samples with rainfall patterns.Based on the hierarchical clustering algorithm and time dynamic adjustment algorithm of data mining,an integrated data-driven model coupled with clustering algorithm is constructed.Through the classification of rainfall events and reconstruction of data sets,the balanced distribution of sample characteristics is realized,so that the data-driven model can learn“information”more accurately,thus improving its runoff prediction ability.Finally,the prediction results of single data driven model and integrated data driven model are compared in the test set.The results showed that the optimal values of NSE,RMSE and R~2 appear in the integrated models and the prediction effect of DTW-LSTM is the most stable.In the test set,except that the NSE of R41 rainfall event is less than 0.8,other rainfall events are greater than 0.8.