Real-time Rainfall-runoff Prediction Of Urban Based On SSA-LightGBM

Timely and accurate prediction of urban rainfall and runoff,so as to formulate reasonable measures for waterlogging prevention and control,is of great significance to maintain normal social and economic order and protect people’s life and property safety.However,because the urban surface runoff is affected by many factors,such as geographical landform,underlying surface,rainfall and the characteristic of rain,it is very difficult to realize the real-time and accurate prediction.Rainfall-runoff physical prediction models are not only complicated in modeling conditions and unstable in forecasting performance,but also difficult to solve.Rainfall-runoff data-driven model based on intelligent algorithm has strong advantages in solving high-strength nonlinear runoff time series problems,and has gradually become a hot spot in basin simulation research in recent years.In this paper,a data-driven model based on SSA-LightGBM architecture was constructed to study the real-time prediction simulation of regional rainfall and runoff by taking a block in Yuelai New Town,Liangjiang New Area,Chongqing City as the object.And the performance of SSA-LightGBM model was verified by analyzing and comparing the prediction effects with several data-driven models commonly used in runoff prediction,such as LSTM model,random forest model and Lightweight Gradient Lift model.The main research contents and achievements are as follows:(1)The hydrological characteristics of runoff series in the study area were analyzed.The results showed that when the lag was 1,the average autocorrelation coefficient of runoff series was 0.861,which had a strong positive correlation.When the lag increased to 6,the autocorrelation coefficient of runoff series was between 0.2 and0.4,showing a weak correlation.The analysis results of single runoff events showed that the runoff sequence was left-skewed.Among the selected 39 runoff events,the runoff events with peak types of single peak,double peak and multi-peak accounted for47%,38% and 15%,respectively,and multi-peak runoff events accounted for the least proportion.(2)A rainfall-runoff data-driven coupling model based on SSA-LightGBM architecture was constructed.In view of the deficiency of Light GBM model in processing strongly nonlinear series data,singular spectrum analysis(SSA)was used to decompose and reconstruct runoff series data,and the trend characteristics,periodicity characteristics and noise components of runoff series were extracted.Further,by eliminating noise and reconstructing two kinds of sequence data with different characteristics,the SSA-LightGBM model exhibited stronger specificity and sensitivity.(3)An automatic optimization program was designed to solve the problem that the model could not automatically update the super parameters with the increase of data in the real-time runoff prediction.The hyperparameters of LSTM,RF,Light GBM and SSA-LightGBM models were optimized by using TPE optimization algorithm based on Bayesian theory.The results showed that the optimized hyperparameters could make the model give full play to the best performance.(4)The SSA-LightGBM coupling model was used to simulate the rainfall and runoff in the study area,and compared with four data-driven models commonly used in runoff prediction.The results showed that the prediction accuracy of LSTM increases with the increase of input step size.RF,Light GBM and SSA-LightGBM have the best prediction accuracy under specific input step size.Therefore,the model performance should be tested experimentally in practical application to determine the best input step size value.When the input step was the best and the time step was different,the maximum NSE coefficient of SSA-LightGBM could reach 0.941.Compared with LSTM,RF and Light GBM,it increased by 11.8%,7.6% and 7.5%,respectively.In terms of training and test time,the SSA-LightGBM coupling model consumed less than1/13 of the LSTM model,which could still well meet the requirement of real-time prediction for calculation time.Under a single runoff event,the overall prediction performance and peak flow prediction of SSA-Light GBM coupling model are better than LSTM,RF and Light GBM models,and can be applied to the real-time prediction of urban rainfall and runoff.