Research On Flash Flood Forecasting Based On Long Short-Term Memory Networks

Flash floods are among the most destructive natural disasters in many countries of the world.Flash flood prevention is not only a major issue facing the world,but also a tough challenge for China.In China,flash floods are characterized by widespread distribution,large quantities,high burstiness,and destructiveness attributed to complex geographic and geomorphic conditions and various climate types.Flash flood forecasting and early-warning can provide the necessary scientific support for flash flood prevention.It is a kernel study of flash flood forecasting and warning to establish a practical real-time forecasting model with high prediction accuracy,strong application ability,simple calculation and easy popularization,which is applicable to small and medium catchments in hilly areas.This thesis aims to solve key problems and technical difficulties of flash flood forecasting and early-warning and its interdisciplinary research with deep learning.Taking Anhe catchment in Jiangxi province of China as a case study,this thesis studies on the modeling approach,uncertainty analysis and interpretability of machine learning flash flood forecasting modeling based on Long-short term memory(LSTM)and stack architecture of deep learning.Under the background of both imperious demands of flash flood forecasting and early-warning and explosive development of deep learning,the research work of this thesis has great theoretical significance and engineering application value for improving the innovation ability of flash flood forecasting and early-warning methods and the application of artificial intelligence technology.The main contents and innovative achievements are as follows:(1)With the purpose of constructing machine learning flash flood forecasting models which are applicable to small and medium catchments in hilly areas,the hydrological characteristics of Anhe catchment are analyzed.Firstly,the general situation of Anhe catchment is introduced,and the hydrological data is checked and compiled.Secondly,typical flood events are selected and used to identify the characteristics of rainfall and flood in small and medium hilly catchments.Thirdly,three sample sets are constructe,which are applicable to the evaluation of machine learning models,the uncertainty analysis of machine learning modeling,the simulation of both Xin 'anjiang model and machine learning models,respectively.The flood evaluation indexes and standard are determined,which are in line with the practical requirements of flash flood forecasting and early-warning.Data foundation and modeling preparation are provided for machine learning flash flood forecasting modeling,weight visualization analysis,modeling uncertainty analysis,feature impact analysis and hydrologic interpretation of LSTM memory cell.(2)In order to solve the problem that it is difficult for conventional hydrological models to achieve forecast accuracy of short lead-time in small and medium hilly catchments,this thesis proposes a machine learning flash flood forecasting modeling approach and the weight visualization analysis method.Firstly,single-step LSTM and single-output BP flood forecasting models are established based on LSTM,BP neural network and stack architecture.Secondly,the forecast accuracy and generalization ability of these two models are compared and analyzed.Two different forecasting modes are designed to demonstrate the practicability of single-step LSTM flood forecasting models.Thirdly,weight visualization analysis is carried out to vertify that the internal calculation of single-step LSTM flood forecasting models conforms to the time sequence characteristics of the hydrological calculation.This approach integrates the analysis of hydrological characteristics into the modeling process of the machine learning flood forecast model,and realizes the high-precision prediction of short lead-time in small and medium hilly catchments.(3)Aiming at the problem that the current uncertainty analysis of machine learning flood forecasting model only takes single modeling factor impact into consideration,this thesis proposes a framework for uncertainty quantification in machine learning modeling for multi-step time-series forecasting and evaluation of machine learning approach.Firstly,three main uncertainty sources of machine learning modeling are expounded and the framework is proposed based on the analysis of variance theory.Secondly,the simulation results of each combination scheme in different lead-time are calculated to identify the advantages of LSTM in machine learning flood forecasting modeling with multiple uncertainty sources.Thirdly,the uncertainty contribution of the sample set,machine learning approach,deep learning architecture and their interactions are quantified in different lead-time and different discharge quantiles.The proposed framework analyzes the influence of several modeling factors on the flash flood forecasting results with the change of lead-time and discharge quantiles,and fills the gap that the evaluation of machine learning cannot take into account other modeling factors.(4)To address the problem that it is difficult for the current interpretability analysis method to verify the rationality of machine learning flood forecasting model,this thesis presents a method of feature impact analysis for machine learning flood forecasting model.Firstly,multi-step LSTM flood forecasting model is established on the basis of the uncertainty analysis of machine learning flood forecasting modeling.Secondly,the forecast accuracy and generalization ability of single-step and multi-step LSTM flood forecasting model are compared and analyzed at 1?10h leat-time.Thirdly,research approaches are degined for feature impact analysis.The effects of rainfall features,discharge feature and their interaction on the output are separated at each lead-time,so as to analyze the change rule of these features with the flood process.This method reveals the effect of rainfall features keeps pace with rainfall process and the effect of discharge features is consistent with the lead-time of forecast runoff.LSTM multi-step output flood prediction model can excavate the influence law of rainfall spatial distribution on flood process,but it does not follow the principle of water balance in hydrology.Multi-step LSTM flood forecasting model can excavate the influence law of rainfall spatial distribution on flood process,but it it does not follow the principle of water balance in hydrology.(5)For the problem that the model structure and parameters of the machine learning flood forecasting model cannot be interpretated in hydrology,the thesis proposes the method of constructing LSTM conceptual flood forecasting model and its hydrologic interpretation based on deep learning stack architecture.Firstly,the similarities and differences between machine learning model and hydrological model are compared.LSTM conceptual flood forecasting model is established by integrating the structure of the conceptual hydrological model,the advantage of inputting multiple rainfall features and the powerful ability of LSTM to deal with the long-term dependence problem into deep learning stack architectures.Secondly,the forecast accuracy and generalization ability of LSTM conceptual flood forecasting model are compared to Xin'anjiang model and Simple RNN conceptual flood forecasting model in order to evaluate its forecast effect.Thirdly,the state variables of LSTM conceptual flood fotecasting model and Xin 'anjiang model were extracted.Hydrologic interpretation of LSTM memory cell state is revealed.This method breaks through the bottleneck that machine learning flood forecasting model cannot be interpretated in hydrology,and provide new modeling perspective for the research of hydrological simulation and flood forecasting.Combined discharge forecasting realizes the high-precision prediction in small and medium hilly catchments.