Research On Flood Forecasting For Small And Medium-sized Mountainous Rivers Based On Long Short-term Memory Networks

Due to the extreme climate and human activities,local storm rainfall events occur frequently in mountainous areas,and flash flood disaster has become more serious over past several years.Flash flood is also the main natural disasters causing much human death in China,which brings a large barrier to socio-economic development and rural revitalization strategy implementation in hilly regions.The flash flood prevention is also a short board and a large difficulty existed in disaster reduction and emergency plan of China.Flood forecasting is an important nonstructural tool used for flood prevention and reduction,and improving the ability of flash flood forecasting and early warning becomes the urgent demands for flash flood prevention in China,which is also a global hot spot and scientific research problem.This paper focuses on solving the problem of high difficulty and low accuracy faced in flood forecasting within mountainous catchments,and machine learning(ML)technique is introduced in this field.Anhe catchment located in Jiangxi province is selected as study site.This paper concentrates on the key problems and technical difficulties existed in research on the crossapplication of machine learning technique and flood forecasting,and the following three aspects are investigated: the development of machine learning model used for flash flood forecasting,the determination of machine learning flood forecasting model’s structures and parameters,and the interpretability of machine learning flood forecasting model.The improvement of the flood forecasting level in mountainous catchment is expected by using new techniques,which also provides technical support for flash flood prevention in China.The main research contents and achievements of this paper are as follows:(1)Conceptual hydrological model is now the most important flood forecasting tool,and is also the most popular hydrological model in the practical usage of flood forecasting.A conceptual hydrological model(i.e.,Xin’anjiang(XAJ)model)is selected as the benchmark model and constructed for flash flood forecasting.Firstly,the basic principle,model structure and parameters of XAJ conceptual hydrological model are briefly introduced.Then,the parameter optimization of XAJ model and relevant optimization objective functions are described in detail.Finally,the performance of XAJ model in the selected mountainous catchment is analyzed and discussed.The results show that the simulated and predicted accuracy of XAJ conceptual hydrological model is acceptable in Anhe catchment,and the qualified rate of simulated and predicted peak flows during calibration and testing periods is0.757 and 0.70,respectively,and the average Nash efficiency coefficient(NSE)during calibration and testing periods is 0.779 and 0.71,respectively.(2)In order to solve the difficulty in flash flood forecasting and early warning,and the problems existed in the traditional hydrological model with respect to the complex modeling,high requirements for hydrological measured data and low prediction accuracy,this paper introduces the machine learning technique and relevant data pre-and post-processing methods.The Long-Term Memory(LSTM)networks is selected as the flood forecasting tool,and a flood forecasting method that is suitable for small and medium-sized catchments in hilly areas has been proposed using machine learning technique.Firstly,the basic principle and internal calculation mechanism of LSTM model are briefly introduced.Then,the structure of LSTM flood forecasting model is introduced in detail,and the structures and their characteristics are analyzed from the following aspects: the model input and output relationship,the number of hidden layers and the number of hidden neuron nodes.Next,the parameters of LSTM flood forecasting model and their main factors that influence the determination of model parameters are described in detail,in which the associated mechanism of hyper-parameters,loss function,optimization algorithm and activation function on the determination of model parameters are explored.Finally,taking the training process of LSTM model as the core,the hierarchical relationship among the model structure,parameter and its impact factors is described in detail.(3)Taking Anhe catchment as an example,an instantiated flood forecasting model used for mountainous catchment is established based on the proposed method using LSTM,and its practicability in flash flood forecasting is investigated.Firstly,the relationship between model inputs and outputs is firstly determined according to the confluence time of the study catchment.A multi-step LSTM flood forecasting model fed with a long sequence of stational rainfall data is constructed.In order to analyze and investigate the characteristics and performance of LSTM flood forecasting model,different model structures as well as the combination scheme of external factors affecting the determination of model parameters are designed.Then,the simulated and predicted results from LSTM model in the study catchment are extracted and analyzed,and compared with the traditional XAJ conceptual hydrological model.Finally,the performance of LSTM model for real-time flood forecasting is compared and analyzed.The results show that the LSTM model has strong fitting and generalization capability,and that the qualified rates of simulated and predicted peak flows in Anhe catchment can reach more than0.80,and the calculated average NSE values of the simulated and predicted flood hydrographs are 0.871 and 0.821 respectively.Moreover,the performance of LSTM model in real-time flood forecasting is consistent with the XAJ model,which can effectively assist in flood forecasting and early warning in mountainous catchment.LSTM model can also achieve acceptable performance in another mountainous catchment located in northeastern China.(4)In order to solve the problem that the forecasting models based on machine learning techniques are generally lack of interpretability,the constructed LSTM model is compared and analyzed with XAJ conceptual hydrological model in this dissertation.The internal learning mechanism of LSTM model is analyzed and explored from the following aspects: the internal structure characteristics of the LSTM model,the information hidden in input and output data and the calculation flow chart existed in LSTM,long-term and short-term memory mechanism,temporal variation of internal state time series from LSTM model and their relationships with XAJ model parameters or intermediate variables.The hydrological interpretation of LSTM model is provided from the perspective of the implicit relationship of intermediate variables,and finally realizes the qualitative and quantitative analysis and interpretation of LSTM model.The results show that the built LSTM model takes fully advantage of its long-term memory mechanism,and converts the long sequence of input rainfall information into the internal states of model control gates and cell states,meanwhile,propagates them forward along the time sequence.There is a strong positive or negative correlation existed between the internal state variables of the LSTM model and the input rainfall or target discharge values.The control gate and cell state of LSTM model essentially remember the temporal variation characteristics of the sequence of input rainfall and target discharge information,so as to establish the internal conversion relationship from rainfall to discharge at the outlet of catchment.The internal state variables of LSTM flood forecasting model are also significantly related to the parameters and intermediate calculation variables of XAJ model.In particular,the hidden cell state of LSTM and the free water storage of XAJ have large similarity from the perspective of physical and statistical analysis.